1
00:00:00,000 --> 00:00:02,440
Your agent starts fast by month three at sluggish.

2
00:00:02,440 --> 00:00:04,360
You notice it first in the response times.

3
00:00:04,360 --> 00:00:06,800
A question that took three seconds now takes eight.

4
00:00:06,800 --> 00:00:08,280
You add more context hoping it helps.

5
00:00:08,280 --> 00:00:11,640
You paste in policies, procedures, guardrails, and examples.

6
00:00:11,640 --> 00:00:14,040
You want the model to have everything it might need.

7
00:00:14,040 --> 00:00:16,080
But then the cost goes up, the errors increase,

8
00:00:16,080 --> 00:00:17,120
and the agent gets slower.

9
00:00:17,120 --> 00:00:18,480
So you do what seems logical.

10
00:00:18,480 --> 00:00:19,560
You add a bigger model.

11
00:00:19,560 --> 00:00:20,600
You refine the prompt.

12
00:00:20,600 --> 00:00:22,400
You try different orchestration patterns.

13
00:00:22,400 --> 00:00:23,440
None of it works.

14
00:00:23,440 --> 00:00:25,960
Because the problem isn't the model, the problem is the structure.

15
00:00:25,960 --> 00:00:28,280
Every turn that agent reads the entire manual,

16
00:00:28,280 --> 00:00:31,040
every token in your system prompt gets processed.

17
00:00:31,040 --> 00:00:33,920
Every policy, every procedure, and every edge case

18
00:00:33,920 --> 00:00:35,400
is something the model attends to,

19
00:00:35,400 --> 00:00:36,800
whether it's relevant or not.

20
00:00:36,800 --> 00:00:38,080
That's the context text.

21
00:00:38,080 --> 00:00:40,000
And it's killing enterprise AI deployments.

22
00:00:40,000 --> 00:00:41,880
The fix isn't bigger models or better prompts.

23
00:00:41,880 --> 00:00:42,960
It's architecture.

24
00:00:42,960 --> 00:00:44,480
Why agents degrade?

25
00:00:44,480 --> 00:00:46,400
This is what most organizations don't understand

26
00:00:46,400 --> 00:00:47,760
about agent decay.

27
00:00:47,760 --> 00:00:50,280
It starts with what feels like a simple design choice.

28
00:00:50,280 --> 00:00:52,400
You write a comprehensive instruction set.

29
00:00:52,400 --> 00:00:54,320
You document how the agent should behave,

30
00:00:54,320 --> 00:00:57,160
what tone it should adopt, and what rules it must follow.

31
00:00:57,160 --> 00:00:59,200
You add domain knowledge like your data model,

32
00:00:59,200 --> 00:01:01,400
your business rules, and your compliance requirements.

33
00:01:01,400 --> 00:01:02,320
You throw in examples.

34
00:01:02,320 --> 00:01:03,560
You anticipate edge cases.

35
00:01:03,560 --> 00:01:05,560
The result is a monolithic instruction blob.

36
00:01:05,560 --> 00:01:09,720
Everything in one place, one system prompt, one version, one owner.

37
00:01:09,720 --> 00:01:11,280
For the first few weeks, it works fine.

38
00:01:11,280 --> 00:01:12,320
The agent is fast.

39
00:01:12,320 --> 00:01:13,200
The responses are good.

40
00:01:13,200 --> 00:01:14,400
Users adopt it.

41
00:01:14,400 --> 00:01:15,920
Then you need to add something new.

42
00:01:15,920 --> 00:01:16,840
A new policy.

43
00:01:16,840 --> 00:01:19,000
A different workflow for a specific department.

44
00:01:19,000 --> 00:01:21,720
A guardrail for a recent incident, you add it to the blob.

45
00:01:21,720 --> 00:01:22,840
The blob grows.

46
00:01:22,840 --> 00:01:24,640
By month three, that instruction set

47
00:01:24,640 --> 00:01:27,480
has become this massive, undifferentiated pile of rules.

48
00:01:27,480 --> 00:01:28,680
Some are critical.

49
00:01:28,680 --> 00:01:30,720
Some apply only once a quarter.

50
00:01:30,720 --> 00:01:32,640
Some contradict each other in subtle ways.

51
00:01:32,640 --> 00:01:34,280
But the agent loads all of it.

52
00:01:34,280 --> 00:01:35,280
Every single turn.

53
00:01:35,280 --> 00:01:36,680
That's context saturation.

54
00:01:36,680 --> 00:01:37,920
Here's what happens at scale.

55
00:01:37,920 --> 00:01:39,360
The model reads more tokens.

56
00:01:39,360 --> 00:01:40,840
Reading more tokens costs more.

57
00:01:40,840 --> 00:01:42,000
It also takes longer.

58
00:01:42,000 --> 00:01:43,240
Latency increases.

59
00:01:43,240 --> 00:01:44,720
The reasoning parts gets noisier

60
00:01:44,720 --> 00:01:47,120
because the model has to pass through irrelevant guidance

61
00:01:47,120 --> 00:01:48,360
to find what matters.

62
00:01:48,360 --> 00:01:49,560
So accuracy drifts.

63
00:01:49,560 --> 00:01:50,640
False positives increase.

64
00:01:50,640 --> 00:01:53,480
The agent starts making mistakes it didn't make before.

65
00:01:53,480 --> 00:01:55,000
You notice the agent feels dumb.

66
00:01:55,000 --> 00:01:55,840
It is.

67
00:01:55,840 --> 00:01:58,440
Not because the model got worse, but because it's drowning in noise.

68
00:01:58,440 --> 00:01:59,760
The research is clear on this.

69
00:01:59,760 --> 00:02:01,560
When you split a monolithic instruction set

70
00:02:01,560 --> 00:02:04,480
into modular skills, loading only the guidance relevant

71
00:02:04,480 --> 00:02:05,520
to the current task.

72
00:02:05,520 --> 00:02:08,800
Organizations report 96% token savings.

73
00:02:08,800 --> 00:02:11,000
Not a small percentage, 96%.

74
00:02:11,000 --> 00:02:14,240
That same split delivers a 93% latency improvement.

75
00:02:14,240 --> 00:02:15,360
9/10 faster.

76
00:02:15,360 --> 00:02:17,600
But the issue runs deeper than just speed and cost.

77
00:02:17,600 --> 00:02:19,360
There's a structural flaw nobody talks about.

78
00:02:19,360 --> 00:02:21,280
Organizations treat all rules the same way.

79
00:02:21,280 --> 00:02:24,560
A universal guardrail that must apply to every conversation

80
00:02:24,560 --> 00:02:26,920
gets piled into the same instruction file

81
00:02:26,920 --> 00:02:28,840
as a specialized workflow that applies

82
00:02:28,840 --> 00:02:31,520
to maybe 5% of your use cases.

83
00:02:31,520 --> 00:02:33,240
The specialized guidance is always loaded,

84
00:02:33,240 --> 00:02:35,560
always consuming tokens, and always slowing down

85
00:02:35,560 --> 00:02:36,320
the reasoning path.

86
00:02:36,320 --> 00:02:38,160
Even though most of the time it's irrelevant.

87
00:02:38,160 --> 00:02:39,960
The difference is between always on guidance

88
00:02:39,960 --> 00:02:41,360
and situational expertise.

89
00:02:41,360 --> 00:02:45,320
Always on guidance is your tone, your role, your identity.

90
00:02:45,320 --> 00:02:47,200
Universal compliance rules.

91
00:02:47,200 --> 00:02:49,360
Things that apply to every single conversation,

92
00:02:49,360 --> 00:02:51,320
regardless of what the user asks for.

93
00:02:51,320 --> 00:02:53,120
Situational expertise is different.

94
00:02:53,120 --> 00:02:55,280
It's the step-by-step procedure for handling a refund

95
00:02:55,280 --> 00:02:56,680
above a certain threshold.

96
00:02:56,680 --> 00:02:58,760
It's the checklist for onboarding a new vendor.

97
00:02:58,760 --> 00:03:01,560
It's the specific tax rules for a particular region

98
00:03:01,560 --> 00:03:03,720
invoked only when that region matters.

99
00:03:03,720 --> 00:03:05,920
It's specialized knowledge that the model doesn't need

100
00:03:05,920 --> 00:03:08,400
in context unless the scenario actually comes up.

101
00:03:08,400 --> 00:03:10,960
When you throw both into the same instruction file,

102
00:03:10,960 --> 00:03:12,280
you've created a bloated context

103
00:03:12,280 --> 00:03:14,880
that the agent has to pass on every turn.

104
00:03:14,880 --> 00:03:18,040
The model wastes computation reading irrelevant procedures.

105
00:03:18,040 --> 00:03:20,240
The latency balloons, the cost climbs,

106
00:03:20,240 --> 00:03:22,600
and the accuracy suffers because important guidance

107
00:03:22,600 --> 00:03:23,520
gets buried.

108
00:03:23,520 --> 00:03:24,720
This is the structural flaw.

109
00:03:24,720 --> 00:03:26,240
The moment your agent starts handling

110
00:03:26,240 --> 00:03:27,920
more than one type of scenario,

111
00:03:27,920 --> 00:03:30,680
a monolithic architecture breaks, not immediately,

112
00:03:30,680 --> 00:03:32,360
but predictably, by design.

113
00:03:32,360 --> 00:03:33,840
The fix is to separate them.

114
00:03:33,840 --> 00:03:35,800
Keep the always on guidance lean and sharp.

115
00:03:35,800 --> 00:03:37,640
Move situational guidance into skills.

116
00:03:37,640 --> 00:03:40,480
These are reusable, on-demand instruction sets

117
00:03:40,480 --> 00:03:42,800
that the orchestrator loads only when needed.

118
00:03:42,800 --> 00:03:45,320
Load the one thing the agent actually needs for this task.

119
00:03:45,320 --> 00:03:46,880
Ignore the other 47 draws.

120
00:03:46,880 --> 00:03:47,960
That's the architecture shift.

121
00:03:47,960 --> 00:03:49,400
That's what modern agents are doing.

122
00:03:49,400 --> 00:03:51,320
And it's where we're heading next.

123
00:03:51,320 --> 00:03:53,120
The skill as a reasoning boundary.

124
00:03:53,120 --> 00:03:54,600
So what exactly is a skill?

125
00:03:54,600 --> 00:03:56,960
At its core, a skill is a self-contained package.

126
00:03:56,960 --> 00:03:58,920
It contains three things, structured instructions

127
00:03:58,920 --> 00:04:00,680
for a specific workflow, metadata

128
00:04:00,680 --> 00:04:02,720
that tells the orchestrator when to use it,

129
00:04:02,720 --> 00:04:03,960
and bundled resources.

130
00:04:03,960 --> 00:04:05,760
Like scripts, templates, or documents,

131
00:04:05,760 --> 00:04:07,160
the foundation is the skill.

132
00:04:07,160 --> 00:04:09,520
MD file, this is just a markdown document

133
00:04:09,520 --> 00:04:10,800
with the YAML front matter.

134
00:04:10,800 --> 00:04:13,480
It declares the name, a description of when it applies

135
00:04:13,480 --> 00:04:14,920
and the procedural steps.

136
00:04:14,920 --> 00:04:17,240
It sounds straightforward, almost boring.

137
00:04:17,240 --> 00:04:19,720
But the format matters because it has become portable.

138
00:04:19,720 --> 00:04:22,600
In late 2025, I'm tropic published an open standard

139
00:04:22,600 --> 00:04:24,960
that created a shared language for these capabilities.

140
00:04:24,960 --> 00:04:27,800
That format was quickly adopted by OpenAI, Microsoft,

141
00:04:27,800 --> 00:04:29,280
and dozens of other platforms.

142
00:04:29,280 --> 00:04:31,600
This didn't happen because of a committee or a mandate.

143
00:04:31,600 --> 00:04:33,080
It happened because it worked.

144
00:04:33,080 --> 00:04:34,760
The format was simple enough to adopt,

145
00:04:34,760 --> 00:04:36,680
expressive enough for real workflows,

146
00:04:36,680 --> 00:04:38,840
and lightweight enough to move between systems

147
00:04:38,840 --> 00:04:40,240
without any translation.

148
00:04:40,240 --> 00:04:42,120
Today, you can write a skill once and use it

149
00:04:42,120 --> 00:04:44,040
across GitHub co-pilot, Claude code,

150
00:04:44,040 --> 00:04:46,880
and co-pilot studio without rewriting a single line.

151
00:04:46,880 --> 00:04:48,480
But here is the part most people miss,

152
00:04:48,480 --> 00:04:50,720
the metadata, the name and the description.

153
00:04:50,720 --> 00:04:52,880
Isn't just documentation, it is an API.

154
00:04:52,880 --> 00:04:53,720
Think about it this way.

155
00:04:53,720 --> 00:04:55,200
The orchestrator has one job.

156
00:04:55,200 --> 00:04:57,760
It reads the user's request, looks at the context,

157
00:04:57,760 --> 00:04:59,240
and asks the single question,

158
00:04:59,240 --> 00:05:01,240
which skill should I load right now?

159
00:05:01,240 --> 00:05:03,120
To answer that, the orchestrator does not read

160
00:05:03,120 --> 00:05:04,200
the full instructions.

161
00:05:04,200 --> 00:05:07,080
It doesn't pass 30 paragraphs of step-by-step guidance.

162
00:05:07,080 --> 00:05:09,200
Instead, it evaluates the name and description

163
00:05:09,200 --> 00:05:11,200
against what the user is trying to do.

164
00:05:11,200 --> 00:05:13,160
The name and description are the routing signals.

165
00:05:13,160 --> 00:05:15,520
This is why precision in naming is everything.

166
00:05:15,520 --> 00:05:17,400
If you name a skill HR help, it will fail

167
00:05:17,400 --> 00:05:18,960
because the scope is too broad.

168
00:05:18,960 --> 00:05:21,200
The orchestrator can't tell if the user needs help

169
00:05:21,200 --> 00:05:23,160
with benefits, pay, or leave policies.

170
00:05:23,160 --> 00:05:25,240
The result is that the system either loads the skill

171
00:05:25,240 --> 00:05:28,400
every time HR is mentioned, which pollutes the context window

172
00:05:28,400 --> 00:05:30,760
or it never loads at all because the description is too vague

173
00:05:30,760 --> 00:05:31,560
to match.

174
00:05:31,560 --> 00:05:34,200
But HR leave eligibility triage actually works.

175
00:05:34,200 --> 00:05:36,520
The name is specific, the description is precise.

176
00:05:36,520 --> 00:05:38,320
You tell the system to use this skill

177
00:05:38,320 --> 00:05:40,680
when a user asks about leave eligibility,

178
00:05:40,680 --> 00:05:42,960
require documents or a crew calculations.

179
00:05:42,960 --> 00:05:45,120
Now the orchestrator can make a real decision.

180
00:05:45,120 --> 00:05:47,120
When a user asks how much paid time off they have,

181
00:05:47,120 --> 00:05:49,080
the model reads that description, sees the match,

182
00:05:49,080 --> 00:05:50,080
and loads the skill.

183
00:05:50,080 --> 00:05:52,040
But when that same user asks about their salary,

184
00:05:52,040 --> 00:05:53,480
the skill stays unloaded.

185
00:05:53,480 --> 00:05:55,880
The orchestrator's job is matching intent to expertise.

186
00:05:55,880 --> 00:05:58,360
Once the skill is selected, the full instructions

187
00:05:58,360 --> 00:05:59,840
finally come into context.

188
00:05:59,840 --> 00:06:01,720
Now the model has the step-by-step procedure,

189
00:06:01,720 --> 00:06:04,480
the examples, and the guardrails for that specific workflow.

190
00:06:04,480 --> 00:06:05,760
It executes with precision.

191
00:06:05,760 --> 00:06:07,840
And when the conversation shifts, the orchestrator

192
00:06:07,840 --> 00:06:09,680
unloads it and brings in something else.

193
00:06:09,680 --> 00:06:11,040
This is the reasoning boundary.

194
00:06:11,040 --> 00:06:13,120
A skill isn't a straight jacket that forces the model

195
00:06:13,120 --> 00:06:14,360
to act a certain way.

196
00:06:14,360 --> 00:06:17,040
It is a boundary that marks where one area of expertise

197
00:06:17,040 --> 00:06:18,640
ends and another begins.

198
00:06:18,640 --> 00:06:21,080
It tells the model that the specialized knowledge applies

199
00:06:21,080 --> 00:06:22,920
to this specific scenario, but it

200
00:06:22,920 --> 00:06:25,040
shouldn't treat this guidance as universal.

201
00:06:25,040 --> 00:06:27,240
Skills also use what we call soft-pointing.

202
00:06:27,240 --> 00:06:29,320
This means they reference tools without being permanently

203
00:06:29,320 --> 00:06:30,320
bound to them.

204
00:06:30,320 --> 00:06:33,560
A skill can tell the model to call a leave-approval action,

205
00:06:33,560 --> 00:06:36,160
but it doesn't require that action to exist to function.

206
00:06:36,160 --> 00:06:38,960
If the action isn't in the toolkit, the skill still loads,

207
00:06:38,960 --> 00:06:41,760
and the model simply adapts by acknowledging the limitation

208
00:06:41,760 --> 00:06:43,360
or escalating the issue.

209
00:06:43,360 --> 00:06:46,240
The skill provides the blueprint, not an iron-clad mandate.

210
00:06:46,240 --> 00:06:48,640
This architecture separates instruction from execution

211
00:06:48,640 --> 00:06:50,040
and knowledge from tools.

212
00:06:50,040 --> 00:06:51,680
Each piece can evolve on its own.

213
00:06:51,680 --> 00:06:53,920
And the orchestrator, the intelligent router,

214
00:06:53,920 --> 00:06:55,840
decides what gets loaded and when.

215
00:06:55,840 --> 00:06:57,000
That is the shift.

216
00:06:57,000 --> 00:06:58,920
That is what a skill actually is.

217
00:06:58,920 --> 00:07:00,560
The progressive disclosure pattern.

218
00:07:00,560 --> 00:07:02,920
There is a design principle underneath all of this

219
00:07:02,920 --> 00:07:04,520
called progressive disclosure.

220
00:07:04,520 --> 00:07:05,560
It works like this.

221
00:07:05,560 --> 00:07:07,680
Your agent does not load the entire skill library

222
00:07:07,680 --> 00:07:08,440
at startup.

223
00:07:08,440 --> 00:07:11,320
Instead, it loads the metadata, the lightweight summary.

224
00:07:11,320 --> 00:07:14,000
For every skill, that is just the name and the description.

225
00:07:14,000 --> 00:07:16,120
It is compact and uses minimal tokens.

226
00:07:16,120 --> 00:07:18,080
Think of it as a directory that the orchestrator keeps

227
00:07:18,080 --> 00:07:20,400
in view at all times to decide what is relevant.

228
00:07:20,400 --> 00:07:22,640
Everything else, the instructions, the examples,

229
00:07:22,640 --> 00:07:26,080
and the scripts stays on the disk until the moment it is needed.

230
00:07:26,080 --> 00:07:27,920
When the orchestrator decides, a skill

231
00:07:27,920 --> 00:07:30,880
matches the user's intent, the full content loads.

232
00:07:30,880 --> 00:07:33,040
The instructions enter the context mid-conversation,

233
00:07:33,040 --> 00:07:34,200
the reasoning path shifts.

234
00:07:34,200 --> 00:07:36,440
Suddenly, the model has access to specialized guidance

235
00:07:36,440 --> 00:07:38,560
that was not available a few seconds ago.

236
00:07:38,560 --> 00:07:39,840
This matters for operations in ways

237
00:07:39,840 --> 00:07:41,320
that go away beyond the obvious.

238
00:07:41,320 --> 00:07:43,320
When a skill is unloaded, those tokens

239
00:07:43,320 --> 00:07:44,640
are not sitting in the prompt.

240
00:07:44,640 --> 00:07:46,000
They aren't there at all.

241
00:07:46,000 --> 00:07:47,680
This means that for every response,

242
00:07:47,680 --> 00:07:49,520
the model processes a smaller prompt

243
00:07:49,520 --> 00:07:51,400
with fewer tokens to attend to.

244
00:07:51,400 --> 00:07:53,120
Processing time drops as a result.

245
00:07:53,120 --> 00:07:55,520
A conversation that would have taken 45 seconds

246
00:07:55,520 --> 00:07:58,800
with a bloated monolithic prompt might only take 15 seconds

247
00:07:58,800 --> 00:08:00,720
when only the active skill is loaded.

248
00:08:00,720 --> 00:08:02,400
That isn't just a small improvement.

249
00:08:02,400 --> 00:08:03,680
It is a structural change.

250
00:08:03,680 --> 00:08:05,480
The latency math is very simple.

251
00:08:05,480 --> 00:08:07,840
Model inference time scales with token count.

252
00:08:07,840 --> 00:08:09,480
If you cut the tokens per turn and half,

253
00:08:09,480 --> 00:08:10,920
you cut the latency roughly in half.

254
00:08:10,920 --> 00:08:13,600
If you reduce tokens by 80%, which is realistic

255
00:08:13,600 --> 00:08:15,040
when you split situational guidance

256
00:08:15,040 --> 00:08:16,880
out of the main instructions.

257
00:08:16,880 --> 00:08:17,760
The speed follows.

258
00:08:17,760 --> 00:08:19,560
Your response time improves proportionally,

259
00:08:19,560 --> 00:08:23,400
but the real leverage shows up when the system is under load.

260
00:08:23,400 --> 00:08:25,800
When you have multiple conversations happening at once,

261
00:08:25,800 --> 00:08:27,400
throughput becomes the bottleneck.

262
00:08:27,400 --> 00:08:29,720
Every conversation consumes compute.

263
00:08:29,720 --> 00:08:31,720
If your instruction set forces every turn

264
00:08:31,720 --> 00:08:33,960
to process 10,000 tokens, you are burning

265
00:08:33,960 --> 00:08:36,640
through your computational budget at an incredible rate.

266
00:08:36,640 --> 00:08:38,160
If you add 50 concurrent users,

267
00:08:38,160 --> 00:08:40,600
you will hit rate limits and start queuing requests.

268
00:08:40,600 --> 00:08:42,160
The service feels sluggish to everyone.

269
00:08:42,160 --> 00:08:44,440
Progressive disclosure changes that equation.

270
00:08:44,440 --> 00:08:45,920
Because skills load on demand,

271
00:08:45,920 --> 00:08:47,880
the baseline context stays lean.

272
00:08:47,880 --> 00:08:49,840
Most turns process a fraction of the tokens

273
00:08:49,840 --> 00:08:52,160
they would in an old school architecture.

274
00:08:52,160 --> 00:08:55,040
50 concurrent users now occupy much less compute

275
00:08:55,040 --> 00:08:56,880
allowing your throughput to scale.

276
00:08:56,880 --> 00:08:58,720
You can handle more conversations simultaneously

277
00:08:58,720 --> 00:09:00,440
before you ever hit a limit.

278
00:09:00,440 --> 00:09:02,320
This is why a skills architecture scales better

279
00:09:02,320 --> 00:09:03,640
than a monolithic agent.

280
00:09:03,640 --> 00:09:05,120
It isn't just that skills are faster.

281
00:09:05,120 --> 00:09:07,880
It is that they let you do more work with the same budget.

282
00:09:07,880 --> 00:09:09,440
The cost lever works the same way.

283
00:09:09,440 --> 00:09:11,120
Every time the model generates a response,

284
00:09:11,120 --> 00:09:12,360
you pay for tokens.

285
00:09:12,360 --> 00:09:13,800
Every search or tool invocation

286
00:09:13,800 --> 00:09:15,840
has a cost measured in co-pilot credits.

287
00:09:15,840 --> 00:09:17,560
When you reduce the tokens per turn

288
00:09:17,560 --> 00:09:19,160
through progressive disclosure,

289
00:09:19,160 --> 00:09:21,680
you reduce the reasoning cost at the same time.

290
00:09:21,680 --> 00:09:23,480
Because the skill architecture lets you handle

291
00:09:23,480 --> 00:09:25,480
more conversations per unit of compute,

292
00:09:25,480 --> 00:09:28,120
you reduce the total API calls per conversation.

293
00:09:28,120 --> 00:09:30,040
You have fewer round trips to external systems

294
00:09:30,040 --> 00:09:32,760
and fewer searches because the model has targeted guidance.

295
00:09:32,760 --> 00:09:34,720
Your spend on co-pilot credits goes down

296
00:09:34,720 --> 00:09:37,160
and that difference is measurable and significant.

297
00:09:37,160 --> 00:09:39,560
Organizations are reporting that they keep their metadata

298
00:09:39,560 --> 00:09:41,480
layer, the names and descriptions.

299
00:09:41,480 --> 00:09:43,160
Under a few hundred tokens total,

300
00:09:43,160 --> 00:09:45,520
as conversations branch into different scenarios,

301
00:09:45,520 --> 00:09:46,840
different skills activate.

302
00:09:46,840 --> 00:09:48,960
One conversation might load three separate skills

303
00:09:48,960 --> 00:09:51,720
over its lifetime while another might only load one.

304
00:09:51,720 --> 00:09:54,360
But every single turn, the baseline stays lean.

305
00:09:54,360 --> 00:09:55,880
Only what is needed is in the context.

306
00:09:55,880 --> 00:09:57,400
This is the efficiency of the architecture.

307
00:09:57,400 --> 00:09:59,040
It isn't about one single feature.

308
00:09:59,040 --> 00:10:00,840
It is about the structural principle.

309
00:10:00,840 --> 00:10:02,240
Keep the default minimal.

310
00:10:02,240 --> 00:10:03,720
Load what is needed.

311
00:10:03,720 --> 00:10:04,960
Unload when you are done.

312
00:10:04,960 --> 00:10:05,720
Repeat.

313
00:10:05,720 --> 00:10:07,600
Manageability is a first-order problem.

314
00:10:07,600 --> 00:10:10,000
This is where the real value of skills emerges.

315
00:10:10,000 --> 00:10:12,120
And it's almost never the first thing people talk about.

316
00:10:12,120 --> 00:10:14,040
Performance matters, cost matters.

317
00:10:14,040 --> 00:10:16,560
But what determines whether an AI system actually

318
00:10:16,560 --> 00:10:18,960
survives in an organization is manageability?

319
00:10:18,960 --> 00:10:19,880
Can you maintain it?

320
00:10:19,880 --> 00:10:22,080
Can you update it without breaking something else?

321
00:10:22,080 --> 00:10:22,920
Can you own it?

322
00:10:22,920 --> 00:10:25,920
In a monolithic architecture, you have one instruction file,

323
00:10:25,920 --> 00:10:28,440
one massive document, one version, one review cycle,

324
00:10:28,440 --> 00:10:31,200
one owner, at least theoretically, but in reality,

325
00:10:31,200 --> 00:10:32,600
when everything is interconnected,

326
00:10:32,600 --> 00:10:33,960
ownership becomes fuzzy.

327
00:10:33,960 --> 00:10:35,240
Someone adds a guardrail.

328
00:10:35,240 --> 00:10:36,920
Someone else adds a policy.

329
00:10:36,920 --> 00:10:38,440
A third person refines the tone.

330
00:10:38,440 --> 00:10:41,080
Now you have a file that nobody fully understands,

331
00:10:41,080 --> 00:10:42,360
maintained by committee,

332
00:10:42,360 --> 00:10:44,800
with no clear owner accountable for the whole thing.

333
00:10:44,800 --> 00:10:46,320
When you need to change something,

334
00:10:46,320 --> 00:10:47,760
the review surface is enormous.

335
00:10:47,760 --> 00:10:50,320
You're updating a 5,000 line instruction file

336
00:10:50,320 --> 00:10:51,520
where a change to one workflow

337
00:10:51,520 --> 00:10:53,440
can have ripple effects on unrelated logic.

338
00:10:53,440 --> 00:10:54,840
Everything is entangled.

339
00:10:54,840 --> 00:10:57,080
You don't know what will break, so you test exhaustively.

340
00:10:57,080 --> 00:10:59,080
You get nervous, you move slowly,

341
00:10:59,080 --> 00:11:01,560
or you move fast in things break in production.

342
00:11:01,560 --> 00:11:03,800
Neither is good, with modular skills,

343
00:11:03,800 --> 00:11:05,200
the dynamic reverses.

344
00:11:05,200 --> 00:11:08,880
One skill is one version, one owner, one review cycle.

345
00:11:08,880 --> 00:11:12,080
A skill for leave eligibility is owned by the HR team.

346
00:11:12,080 --> 00:11:13,520
While a skill for vendor onboarding

347
00:11:13,520 --> 00:11:14,920
is owned by procurement,

348
00:11:14,920 --> 00:11:16,360
and a skill for incident escalation

349
00:11:16,360 --> 00:11:17,960
is owned by operations.

350
00:11:17,960 --> 00:11:19,920
Each team maintains their own expertise.

351
00:11:19,920 --> 00:11:21,360
Nobody has to be an expert in everything.

352
00:11:21,360 --> 00:11:23,000
The review surface for any single skill

353
00:11:23,000 --> 00:11:25,760
is small enough that humans can reason about it completely.

354
00:11:25,760 --> 00:11:28,360
A hundred line instruction set for a specific workflow

355
00:11:28,360 --> 00:11:31,320
is something a person can read and review in 15 minutes.

356
00:11:31,320 --> 00:11:33,720
But a 5,000 line monolith takes hours,

357
00:11:33,720 --> 00:11:35,680
change management becomes granular.

358
00:11:35,680 --> 00:11:37,760
You update the leave eligibility skill,

359
00:11:37,760 --> 00:11:40,520
and your only risk breaking leave eligibility workflows.

360
00:11:40,520 --> 00:11:42,320
Incident escalation is unaffected.

361
00:11:42,320 --> 00:11:43,840
Vendor onboarding is unaffected.

362
00:11:43,840 --> 00:11:45,480
You can deploy that skill change to production

363
00:11:45,480 --> 00:11:47,160
on a Thursday without nervousness

364
00:11:47,160 --> 00:11:49,040
because the blast radius is defined.

365
00:11:49,040 --> 00:11:51,120
If something goes wrong, you roll back one skill,

366
00:11:51,120 --> 00:11:52,720
not an entire system.

367
00:11:52,720 --> 00:11:54,200
Version history becomes meaningful.

368
00:11:54,200 --> 00:11:56,000
You can see exactly what changed.

369
00:11:56,000 --> 00:11:57,520
When and why?

370
00:11:57,520 --> 00:12:00,120
You can compare versions and understand the delta.

371
00:12:00,120 --> 00:12:01,840
The operational reality is that teams actually

372
00:12:01,840 --> 00:12:03,560
maintain modular systems better.

373
00:12:03,560 --> 00:12:04,560
This isn't anecdotal.

374
00:12:04,560 --> 00:12:06,680
It's how software engineering works at scale.

375
00:12:06,680 --> 00:12:09,400
Micro-Services 1, monoliths lost.

376
00:12:09,400 --> 00:12:12,200
Not because microservices are always faster or smaller

377
00:12:12,200 --> 00:12:13,480
in total lines of code,

378
00:12:13,480 --> 00:12:15,440
but because humans can reason about them.

379
00:12:15,440 --> 00:12:16,760
A team can own a service.

380
00:12:16,760 --> 00:12:18,920
A team can deploy a service independently.

381
00:12:18,920 --> 00:12:20,560
A team can be paged when that service breaks.

382
00:12:20,560 --> 00:12:23,840
You have accountability, clarity, and speed.

383
00:12:23,840 --> 00:12:25,320
Skills work the same way.

384
00:12:25,320 --> 00:12:26,920
The boundaries matter too.

385
00:12:26,920 --> 00:12:28,880
A skill that serves as a specialist.

386
00:12:28,880 --> 00:12:31,680
Narrow expertise called in occasionally is easy to own.

387
00:12:31,680 --> 00:12:32,600
It has one job.

388
00:12:32,600 --> 00:12:34,160
A skill that serves as a playbook,

389
00:12:34,160 --> 00:12:36,440
the known sequence for a recurring situation,

390
00:12:36,440 --> 00:12:38,000
is similarly bounded.

391
00:12:38,000 --> 00:12:41,320
A playbook skill for quarterly reviews has a clear scope.

392
00:12:41,320 --> 00:12:44,760
The steps, the approvals, the outcomes.

393
00:12:44,760 --> 00:12:47,200
A guardrail skill for compliance is explicit.

394
00:12:47,200 --> 00:12:49,320
These are the rules that must never be violated.

395
00:12:49,320 --> 00:12:51,120
Each archetype has a natural boundary.

396
00:12:51,120 --> 00:12:54,000
When you try to combine them, when you throw a reference manual,

397
00:12:54,000 --> 00:12:55,800
a specialist procedure, a playbook,

398
00:12:55,800 --> 00:12:57,680
and a guardrail all into one skill.

399
00:12:57,680 --> 00:13:00,480
You've recreated the monolithic problem at a smaller scale.

400
00:13:00,480 --> 00:13:01,880
The routing becomes ambiguous.

401
00:13:01,880 --> 00:13:03,400
The ownership becomes unclear.

402
00:13:03,400 --> 00:13:05,040
The review surface expands.

403
00:13:05,040 --> 00:13:07,520
And you've defeated the purpose of modular architecture.

404
00:13:07,520 --> 00:13:10,280
So the first order question isn't performance or cost.

405
00:13:10,280 --> 00:13:12,560
Though those follow, its manageability.

406
00:13:12,560 --> 00:13:13,600
Can we own this?

407
00:13:13,600 --> 00:13:14,640
Can we maintain it?

408
00:13:14,640 --> 00:13:16,040
Can we change it safely?

409
00:13:16,040 --> 00:13:18,440
Skills let you answer yes to all three.

410
00:13:18,440 --> 00:13:21,560
The four skill archetypes, not all skills are the same.

411
00:13:21,560 --> 00:13:23,960
And trying to force different types into a single skill

412
00:13:23,960 --> 00:13:26,120
is where most implementations start to break.

413
00:13:26,120 --> 00:13:28,440
There are four distinct patterns, recognizing

414
00:13:28,440 --> 00:13:31,120
which one your building determines whether routing works.

415
00:13:31,120 --> 00:13:33,640
Whether the orchestrator can invoke it at the right moment.

416
00:13:33,640 --> 00:13:35,520
And whether the model can actually follow it.

417
00:13:35,520 --> 00:13:36,880
The first is the reference manual.

418
00:13:36,880 --> 00:13:40,000
This is a skill that teaches the model something it doesn't know.

419
00:13:40,000 --> 00:13:43,240
Your data model, your schema, your business terminology.

420
00:13:43,240 --> 00:13:46,320
The LLM wasn't trained on your internal customer classification

421
00:13:46,320 --> 00:13:48,480
system or your proprietary process names.

422
00:13:48,480 --> 00:13:50,440
So you documented a reference manual skill

423
00:13:50,440 --> 00:13:52,040
is pure knowledge transfer.

424
00:13:52,040 --> 00:13:54,720
It's usually heavy on structure, tables, definitions,

425
00:13:54,720 --> 00:13:57,720
examples, and light on procedure, it says.

426
00:13:57,720 --> 00:14:00,080
Here's the universe of things that exist in our system

427
00:14:00,080 --> 00:14:01,480
and what they mean.

428
00:14:01,480 --> 00:14:04,000
When a user asks a question that requires understanding

429
00:14:04,000 --> 00:14:06,560
your domain, the orchestrator loads this skill.

430
00:14:06,560 --> 00:14:08,120
The model reads the definitions.

431
00:14:08,120 --> 00:14:09,440
And then it can answer accurately.

432
00:14:09,440 --> 00:14:11,920
The reference manual skill doesn't tell the model what to do.

433
00:14:11,920 --> 00:14:13,760
It tells the model what exists.

434
00:14:13,760 --> 00:14:14,960
The second is the specialist.

435
00:14:14,960 --> 00:14:16,760
This is a narrow area of expertise

436
00:14:16,760 --> 00:14:18,960
that the model only occasionally needs.

437
00:14:18,960 --> 00:14:21,480
Think region-specific tax rules applied only

438
00:14:21,480 --> 00:14:23,800
when a user asks about a particular geography

439
00:14:23,800 --> 00:14:27,000
or a specialized approval workflow for high-value orders.

440
00:14:27,000 --> 00:14:29,680
Triggered only when the order amount exceeds a threshold.

441
00:14:29,680 --> 00:14:31,040
The specialist skill is focused.

442
00:14:31,040 --> 00:14:31,880
It has one job.

443
00:14:31,880 --> 00:14:33,360
It's called rarely.

444
00:14:33,360 --> 00:14:35,320
But when it's called, it's called because the situation

445
00:14:35,320 --> 00:14:38,160
is specific and the guidance needs to be precise.

446
00:14:38,160 --> 00:14:39,960
A specialist skill usually contains

447
00:14:39,960 --> 00:14:41,720
decision trees and conditional logic.

448
00:14:41,720 --> 00:14:44,800
If the amount is under 5,000, follow path A.

449
00:14:44,800 --> 00:14:47,480
If it's between 5 and 50,000, follow path B.

450
00:14:47,480 --> 00:14:50,440
If it exceeds 50,000, escalate to this person.

451
00:14:50,440 --> 00:14:53,400
Specialists are easy to root to because their trigger conditions

452
00:14:53,400 --> 00:14:54,160
are clear.

453
00:14:54,160 --> 00:14:55,400
The third is the playbook.

454
00:14:55,400 --> 00:14:57,280
A playbook skill describes a known sequence

455
00:14:57,280 --> 00:14:59,080
for a recurring situation.

456
00:14:59,080 --> 00:15:02,200
On boarding a new customer running a quarterly review cycle,

457
00:15:02,200 --> 00:15:05,200
processing a leave request, a playbook is procedural.

458
00:15:05,200 --> 00:15:06,560
Step one happens.

459
00:15:06,560 --> 00:15:10,480
Then step two, then step three, there are decision points.

460
00:15:10,480 --> 00:15:12,440
But the overall flow is deterministic.

461
00:15:12,440 --> 00:15:14,120
The model's job is to follow the steps,

462
00:15:14,120 --> 00:15:16,200
ask the right questions at each stage,

463
00:15:16,200 --> 00:15:17,880
collect the necessary information,

464
00:15:17,880 --> 00:15:19,320
and move through the sequence.

465
00:15:19,320 --> 00:15:21,200
A playbook skill is narrative in structure.

466
00:15:21,200 --> 00:15:23,840
It reads like instructions because it is instructions.

467
00:15:23,840 --> 00:15:26,440
Most organizational workflows fit this pattern.

468
00:15:26,440 --> 00:15:27,800
The fourth is the guardrail.

469
00:15:27,800 --> 00:15:30,400
A guardrail skill exists to enforce rules that cannot

470
00:15:30,400 --> 00:15:31,160
be violated.

471
00:15:31,160 --> 00:15:33,760
You may not refund an order without going through approval.

472
00:15:33,760 --> 00:15:35,640
You may not grant certain permissions

473
00:15:35,640 --> 00:15:36,960
without audit evidence.

474
00:15:36,960 --> 00:15:39,200
You may not delete records that have legal holds.

475
00:15:39,200 --> 00:15:41,120
A guardrail skill is constraint-based.

476
00:15:41,120 --> 00:15:42,640
It doesn't guide behavior.

477
00:15:42,640 --> 00:15:44,720
It forbids specific behaviors.

478
00:15:44,720 --> 00:15:45,800
It's often negative.

479
00:15:45,800 --> 00:15:48,400
Never do this, rather than positive guidance.

480
00:15:48,400 --> 00:15:50,080
Guardrails are usually brief and explicit

481
00:15:50,080 --> 00:15:52,120
because they're about drawing a hard line.

482
00:15:52,120 --> 00:15:53,360
Not explaining a process.

483
00:15:53,360 --> 00:15:54,440
Here's what matters.

484
00:15:54,440 --> 00:15:57,080
Each archetype has a different triggering condition,

485
00:15:57,080 --> 00:15:59,200
a different structure, and a different relationship

486
00:15:59,200 --> 00:16:00,440
to the model's reasoning.

487
00:16:00,440 --> 00:16:03,160
When the orchestrator evaluates whether to load a skill,

488
00:16:03,160 --> 00:16:06,080
it's matching the user's intent to the skill's scope.

489
00:16:06,080 --> 00:16:07,680
A reference manual skill should load

490
00:16:07,680 --> 00:16:10,360
when the user is asking questions about your data model,

491
00:16:10,360 --> 00:16:12,880
your terminology, or your system structure,

492
00:16:12,880 --> 00:16:16,280
especially skill loads when the triggering condition is met,

493
00:16:16,280 --> 00:16:19,000
like a high-value order or a specific region.

494
00:16:19,000 --> 00:16:21,280
A playbook skill loads when the user wants to initiate

495
00:16:21,280 --> 00:16:23,760
or progress through a known workflow.

496
00:16:23,760 --> 00:16:25,720
A guardrail skill loads when the user is attempting

497
00:16:25,720 --> 00:16:27,080
an action that's restricted.

498
00:16:27,080 --> 00:16:28,240
Now here's where it breaks.

499
00:16:28,240 --> 00:16:30,160
When you mix archetypes in a single skill,

500
00:16:30,160 --> 00:16:32,000
the routing becomes ambiguous.

501
00:16:32,000 --> 00:16:33,320
The description can't be specific

502
00:16:33,320 --> 00:16:35,360
because it's trying to cover multiple patterns.

503
00:16:35,360 --> 00:16:38,360
The name has to be so broad that it matches too many situations.

504
00:16:38,360 --> 00:16:40,880
The orchestrator either loads the skill constantly

505
00:16:40,880 --> 00:16:42,080
or never loads it at all.

506
00:16:42,080 --> 00:16:44,240
And the model's behavior becomes unpredictable.

507
00:16:44,240 --> 00:16:46,560
Is it reading reference material following a procedure

508
00:16:46,560 --> 00:16:47,960
enforcing a constraint?

509
00:16:47,960 --> 00:16:49,960
The skill doesn't have a coherent type.

510
00:16:49,960 --> 00:16:51,920
So the model's reasoning path fragments.

511
00:16:51,920 --> 00:16:53,880
A well-designed skill knows what it is.

512
00:16:53,880 --> 00:16:58,000
Reference manual, specialist, playbook, or guardrail.

513
00:16:58,000 --> 00:17:00,600
Not all four, not even two.

514
00:17:00,600 --> 00:17:02,560
One clear archetype, one clear purpose,

515
00:17:02,560 --> 00:17:06,000
one routing condition that the orchestrator can evaluate precisely.

516
00:17:06,000 --> 00:17:10,560
That clarity is what makes the entire architecture work.

517
00:17:10,560 --> 00:17:12,480
Copilot Studio's skill integration.

518
00:17:12,480 --> 00:17:13,960
This is where theory meets practice.

519
00:17:13,960 --> 00:17:17,160
Copilot Studio implements the agent skills open format directly.

520
00:17:17,160 --> 00:17:19,320
You can author and manage skills within the platform.

521
00:17:19,320 --> 00:17:21,280
No exporting to separate systems.

522
00:17:21,280 --> 00:17:22,760
No learning a different tool set.

523
00:17:22,760 --> 00:17:24,200
The interface is straightforward.

524
00:17:24,200 --> 00:17:26,520
Inside Copilot Studio, there is a skills tab.

525
00:17:26,520 --> 00:17:28,120
From there you have two entry points.

526
00:17:28,120 --> 00:17:29,680
You can create a skill from scratch

527
00:17:29,680 --> 00:17:31,440
or you can upload an existing one.

528
00:17:31,440 --> 00:17:33,880
Creating from scratch gives you a guided workflow.

529
00:17:33,880 --> 00:17:36,680
You provide the name, description, and instructions.

530
00:17:36,680 --> 00:17:39,720
The platform scaffolds the skill.md structure for you.

531
00:17:39,720 --> 00:17:41,520
You fill in the content, you save it,

532
00:17:41,520 --> 00:17:43,520
and the skill is now part of your agent.

533
00:17:43,520 --> 00:17:45,440
Uploading means you already have a skill.

534
00:17:45,440 --> 00:17:46,800
Maybe you authored it elsewhere.

535
00:17:46,800 --> 00:17:49,120
Maybe you downloaded it from a community registry.

536
00:17:49,120 --> 00:17:51,200
Either way, once it is in Copilot Studio,

537
00:17:51,200 --> 00:17:52,640
the platform manages it.

538
00:17:52,640 --> 00:17:55,080
But skills are not just marked on file sitting alone.

539
00:17:55,080 --> 00:17:56,600
They are bundles.

540
00:17:56,600 --> 00:17:59,640
A skill in Copilot Studio can carry bundled resources

541
00:17:59,640 --> 00:18:01,720
along with the skill.md file.

542
00:18:01,720 --> 00:18:04,640
You can include scripts like Python, Shell, or reference code

543
00:18:04,640 --> 00:18:06,840
that the skill uses during execution.

544
00:18:06,840 --> 00:18:09,120
You can bundle references like documentation files,

545
00:18:09,120 --> 00:18:10,520
schemas, and examples.

546
00:18:10,520 --> 00:18:13,160
Anything the agent might read to understand the context

547
00:18:13,160 --> 00:18:14,240
goes in the bundle.

548
00:18:14,240 --> 00:18:17,200
You can include assets like templates, configuration files,

549
00:18:17,200 --> 00:18:18,400
and sample data.

550
00:18:18,400 --> 00:18:20,560
All of this travels together as a single unit.

551
00:18:20,560 --> 00:18:21,600
Think of it like a folder.

552
00:18:21,600 --> 00:18:24,320
The skill.md is the manifest in the instructions.

553
00:18:24,320 --> 00:18:27,200
The script's subdirectory holds your executable code.

554
00:18:27,200 --> 00:18:29,520
References contain the documentation the model reads.

555
00:18:29,520 --> 00:18:31,960
Assets hold your data structures and templates.

556
00:18:31,960 --> 00:18:34,240
When the skill loads, the entire bundle comes with it.

557
00:18:34,240 --> 00:18:36,360
When the skill is versioned, all of these files

558
00:18:36,360 --> 00:18:37,640
are versioned together.

559
00:18:37,640 --> 00:18:39,480
When the skill moves, everything moves.

560
00:18:39,480 --> 00:18:41,520
This bundling approach matters for application

561
00:18:41,520 --> 00:18:42,920
lifecycle management.

562
00:18:42,920 --> 00:18:45,040
Copilot Studio is built on the Power Platform,

563
00:18:45,040 --> 00:18:46,800
which has a mature solution model.

564
00:18:46,800 --> 00:18:49,560
Skills can be packaged into Power Platform Solutions.

565
00:18:49,560 --> 00:18:51,800
That means the skill is not just a loose artifact

566
00:18:51,800 --> 00:18:53,000
sitting in your agent.

567
00:18:53,000 --> 00:18:55,440
It is a managed component that moves through your environment

568
00:18:55,440 --> 00:18:56,240
strategy.

569
00:18:56,240 --> 00:18:58,400
You develop a skill in your development environment.

570
00:18:58,400 --> 00:18:59,320
You test it.

571
00:18:59,320 --> 00:19:01,840
You promote it to a test environment for validation.

572
00:19:01,840 --> 00:19:03,200
Then you promote it to production.

573
00:19:03,200 --> 00:19:05,280
At each stage, the entire bundle moves together

574
00:19:05,280 --> 00:19:06,480
as a versioned unit.

575
00:19:06,480 --> 00:19:08,320
Your pipelines can track skill changes.

576
00:19:08,320 --> 00:19:10,160
Your environment separation is preserved.

577
00:19:10,160 --> 00:19:12,000
Your agent moves with its skills intact.

578
00:19:12,000 --> 00:19:13,600
This is solution-aware ALM.

579
00:19:13,600 --> 00:19:16,200
The skills are not agent-scoped in isolation.

580
00:19:16,200 --> 00:19:18,400
They travel as part of your solution package.

581
00:19:18,400 --> 00:19:20,840
You can update a skill in development, test the change,

582
00:19:20,840 --> 00:19:23,000
and promote it to production without touching other skills

583
00:19:23,000 --> 00:19:24,200
or the agent itself.

584
00:19:24,200 --> 00:19:26,840
Each component has its own version history, its own change

585
00:19:26,840 --> 00:19:29,240
log, and its own deployment cycle.

586
00:19:29,240 --> 00:19:32,000
The soft-pointing pattern is equally important for integration.

587
00:19:32,000 --> 00:19:34,600
A skill can reference agent tools without hard binding

588
00:19:34,600 --> 00:19:35,280
to them.

589
00:19:35,280 --> 00:19:37,680
When a skill says, "Call the approval action,"

590
00:19:37,680 --> 00:19:40,600
it is not declaring that action as a required dependency.

591
00:19:40,600 --> 00:19:42,800
If the action exists in the agent's toolkit,

592
00:19:42,800 --> 00:19:44,160
the skill can invoke it.

593
00:19:44,160 --> 00:19:46,080
If it does not, the instruction can acknowledge

594
00:19:46,080 --> 00:19:47,720
the limitation and work around it.

595
00:19:47,720 --> 00:19:49,080
The agent adapts.

596
00:19:49,080 --> 00:19:51,440
This matters because it means skills can be

597
00:19:51,440 --> 00:19:54,600
authored without assuming a specific agent configuration.

598
00:19:54,600 --> 00:19:56,440
The same skill can work across agents

599
00:19:56,440 --> 00:19:58,280
that have slightly different tool sets

600
00:19:58,280 --> 00:20:00,800
as long as the core capabilities are present.

601
00:20:00,800 --> 00:20:04,320
Soft-pointing extends to flows, connectors, and MCP servers.

602
00:20:04,320 --> 00:20:06,160
A skill can reference a power-automate flow

603
00:20:06,160 --> 00:20:07,360
without binding to it.

604
00:20:07,360 --> 00:20:08,200
The reference is loose.

605
00:20:08,200 --> 00:20:10,680
The model knows the tool exists and tries to use it,

606
00:20:10,680 --> 00:20:12,920
but it can adapt if conditions do not allow.

607
00:20:12,920 --> 00:20:15,480
This flexibility is what enables skills to be reusable

608
00:20:15,480 --> 00:20:18,080
across different environments and agent configurations.

609
00:20:18,080 --> 00:20:20,120
There is one current limitation worth noting.

610
00:20:20,120 --> 00:20:22,360
Skills in Copilot Studio are agent-scoped.

611
00:20:22,360 --> 00:20:24,240
A skill belongs to a specific agent.

612
00:20:24,240 --> 00:20:26,200
You cannot create a tenant-wide skills library

613
00:20:26,200 --> 00:20:28,320
and have all agents consume from it automatically.

614
00:20:28,320 --> 00:20:29,760
That is the roadmap direction.

615
00:20:29,760 --> 00:20:30,880
Shared catalogs.

616
00:20:30,880 --> 00:20:32,320
Cross-product reuse.

617
00:20:32,320 --> 00:20:34,400
A true enterprise skill library.

618
00:20:34,400 --> 00:20:37,400
But today, skills live within the agent they were created for.

619
00:20:37,400 --> 00:20:40,160
If you want to reuse a skill across multiple agents,

620
00:20:40,160 --> 00:20:42,160
you have to import it separately into each one.

621
00:20:42,160 --> 00:20:44,360
That is changing, but it is the current constraint.

622
00:20:44,360 --> 00:20:46,080
The data versus MCP bridge.

623
00:20:46,080 --> 00:20:47,800
Up until now, we have talked about skills

624
00:20:47,800 --> 00:20:49,600
as agent-scoped artifacts.

625
00:20:49,600 --> 00:20:51,320
They live inside a particular agent

626
00:20:51,320 --> 00:20:53,360
and move with it through environments.

627
00:20:53,360 --> 00:20:55,560
But at enterprise scale, this creates a problem.

628
00:20:55,560 --> 00:20:57,720
If you have 50 agents across your organization,

629
00:20:57,720 --> 00:20:59,960
do you replicate the same skill 50 times?

630
00:20:59,960 --> 00:21:01,760
Do you maintain 50 copies of your vendor

631
00:21:01,760 --> 00:21:02,880
on boarding procedure?

632
00:21:02,880 --> 00:21:03,960
That is not scalable.

633
00:21:03,960 --> 00:21:05,000
That is not maintainable.

634
00:21:05,000 --> 00:21:07,440
It is the opposite of what you wanted when you move to skills

635
00:21:07,440 --> 00:21:08,480
in the first place.

636
00:21:08,480 --> 00:21:11,680
This is where the model context protocol enters the architecture.

637
00:21:11,680 --> 00:21:16,040
MCP is an emerging standard that defines how AI agents access tools, data,

638
00:21:16,040 --> 00:21:17,720
and context from external servers.

639
00:21:17,720 --> 00:21:19,640
Multiple vendors are converging on it.

640
00:21:19,640 --> 00:21:22,680
In Copilot Studio, MCP serves a specific purpose.

641
00:21:22,680 --> 00:21:25,040
It lets agents reach out to a centralized server

642
00:21:25,040 --> 00:21:26,880
to discover and invoke skills.

643
00:21:26,880 --> 00:21:28,880
That centralized server is data-verse.

644
00:21:28,880 --> 00:21:32,200
Data-verse is the foundation of Power Apps and Dynamics 365.

645
00:21:32,200 --> 00:21:34,840
Traditionally, it stores operational business data.

646
00:21:34,840 --> 00:21:38,320
Customers orders cases, the transactional entities that run your business.

647
00:21:38,320 --> 00:21:40,840
But in 2026, its role is expanding.

648
00:21:40,840 --> 00:21:42,960
Data-vers is becoming the system of record

649
00:21:42,960 --> 00:21:44,920
for business procedures themselves.

650
00:21:44,920 --> 00:21:47,360
Not just what happened, but how things should happen.

651
00:21:47,360 --> 00:21:52,160
Business skills stored in data-verse are procedures, policies, and playbooks stored as data.

652
00:21:52,160 --> 00:21:56,280
When a vendor on boarding skill is authored, it no longer lives inside a single agent.

653
00:21:56,280 --> 00:21:59,920
It lives in data-verse, indexed, and versioned as a first-class asset.

654
00:21:59,920 --> 00:22:02,520
When a user asks an agent about vendor onboarding,

655
00:22:02,520 --> 00:22:07,000
that agent connects to the data-verse MCP endpoint and queries available skills.

656
00:22:07,000 --> 00:22:09,000
The orchestrator evaluates the user's intent

657
00:22:09,000 --> 00:22:11,360
against skill descriptions stored in data-verse.

658
00:22:11,360 --> 00:22:14,600
On-match, the agent retrieves the skill and loads it into context.

659
00:22:14,600 --> 00:22:16,120
The endpoint is straightforward.

660
00:22:16,120 --> 00:22:19,400
Every data-verse environment has this specific URL structure.

661
00:22:19,400 --> 00:22:22,680
Any agent in the tenant can connect or auth secures the connection.

662
00:22:22,680 --> 00:22:25,840
When an agent connects, it acts as the authenticated user.

663
00:22:25,840 --> 00:22:27,720
Data-vers enforces role-based access.

664
00:22:27,720 --> 00:22:30,120
If a user lacks permission to access certain data,

665
00:22:30,120 --> 00:22:31,640
the agent does not retrieve it.

666
00:22:31,640 --> 00:22:34,040
Security boundaries are preserved transparently.

667
00:22:34,040 --> 00:22:36,320
Cross-environment access becomes possible.

668
00:22:36,320 --> 00:22:39,360
You have separate development and production data-verse environments,

669
00:22:39,360 --> 00:22:42,120
development host, draft skills, and experimental procedures.

670
00:22:42,120 --> 00:22:44,600
Production hosts only certified approved ones.

671
00:22:44,600 --> 00:22:47,760
A development agent connects to development data-verse for testing.

672
00:22:47,760 --> 00:22:50,880
A production agent connects to production data-verse for live use.

673
00:22:50,880 --> 00:22:53,160
The same MCP protocol applies in both.

674
00:22:53,160 --> 00:22:56,440
You do not have different connection patterns or access methods.

675
00:22:56,440 --> 00:22:58,520
It is unified and environment-aware.

676
00:22:58,520 --> 00:23:01,160
The centralization is where governance gains real power.

677
00:23:01,160 --> 00:23:04,040
In an agent-scoped model, a skill is a local artifact

678
00:23:04,040 --> 00:23:05,480
versioned within that agent.

679
00:23:05,480 --> 00:23:09,240
It is reviewed once and it lives or dies tied to that agent's deployment.

680
00:23:09,240 --> 00:23:12,360
But when a skill lives in data-verse, it becomes an enterprise asset.

681
00:23:12,360 --> 00:23:14,680
It is versioned in the native system of data-verse.

682
00:23:14,680 --> 00:23:15,680
It is owned by a team.

683
00:23:15,680 --> 00:23:18,560
It is reviewed once and reused across many agents.

684
00:23:18,560 --> 00:23:20,320
When you update the skill in data-verse,

685
00:23:20,320 --> 00:23:22,960
all consuming agents see the update immediately.

686
00:23:22,960 --> 00:23:23,960
You update once.

687
00:23:23,960 --> 00:23:25,880
50 agents benefit.

688
00:23:25,880 --> 00:23:27,760
The audit trail becomes comprehensive.

689
00:23:27,760 --> 00:23:30,680
Data-verse records who created the skill and when they created it.

690
00:23:30,680 --> 00:23:33,560
It tracks who modified it and when each change occurred.

691
00:23:33,560 --> 00:23:36,560
That trail is queryable, reportable and exportable.

692
00:23:36,560 --> 00:23:40,520
For compliance audits and change management reviews, you have a complete record.

693
00:23:40,520 --> 00:23:42,400
Skills are no longer forensic blind spots.

694
00:23:42,400 --> 00:23:44,320
They are transparent, traced assets.

695
00:23:44,320 --> 00:23:48,600
Skills become organizational intelligence rather than embedded instructions.

696
00:23:48,600 --> 00:23:50,720
They are managed like critical business assets.

697
00:23:50,720 --> 00:23:52,400
They can be evaluated and certified.

698
00:23:52,400 --> 00:23:54,160
They can be assigned clear ownership.

699
00:23:54,160 --> 00:23:57,640
You can mark them as experimental, approved or deprecated.

700
00:23:57,640 --> 00:24:00,520
You can tag them with compliance categories or domain ownership.

701
00:24:00,520 --> 00:24:02,320
The architecture shifts fundamentally.

702
00:24:02,320 --> 00:24:04,360
Instead of agents containing their own knowledge,

703
00:24:04,360 --> 00:24:06,760
agents discover knowledge from a central system.

704
00:24:06,760 --> 00:24:08,360
That shift enables enterprise scale

705
00:24:08,360 --> 00:24:11,200
without sacrificing control or auditability.

706
00:24:11,200 --> 00:24:14,320
A skill created in data-verse becomes the single source of truth.

707
00:24:14,320 --> 00:24:15,760
Multiple agents reference it.

708
00:24:15,760 --> 00:24:16,880
Multiple users benefit.

709
00:24:16,880 --> 00:24:18,880
One governance surface owns it all.

710
00:24:18,880 --> 00:24:20,080
This is the leverage point.

711
00:24:20,080 --> 00:24:23,120
This is where skills move from a feature to a structural principle.

712
00:24:23,120 --> 00:24:26,400
It is the foundation of how enterprise AI operates.

713
00:24:26,400 --> 00:24:28,320
The routing problem solved.

714
00:24:28,320 --> 00:24:31,120
The cleanest theory collapses the moment reality arrives.

715
00:24:31,120 --> 00:24:32,840
You implement skills, you write descriptions,

716
00:24:32,840 --> 00:24:35,800
you deploy your agent, and then something breaks.

717
00:24:35,800 --> 00:24:38,760
A skill fires when it shouldn't, or it never fires at all.

718
00:24:38,760 --> 00:24:41,200
The model invokes the wrong skill for the scenario.

719
00:24:41,200 --> 00:24:44,360
Or it loads multiple conflicting skills simultaneously.

720
00:24:44,360 --> 00:24:45,640
The routing problem is real.

721
00:24:45,640 --> 00:24:47,240
In fact, it's the most common failure mode

722
00:24:47,240 --> 00:24:49,000
in early skill implementations.

723
00:24:49,000 --> 00:24:50,200
But here's the problem.

724
00:24:50,200 --> 00:24:51,760
When a description is vague,

725
00:24:51,760 --> 00:24:54,480
the orchestrator can't make a reliable decision.

726
00:24:54,480 --> 00:24:56,560
Consider a skill description that says,

727
00:24:56,560 --> 00:24:58,840
use this for financial questions.

728
00:24:58,840 --> 00:25:02,240
Now a user asks about the revenue forecast for next quarter.

729
00:25:02,240 --> 00:25:04,520
That matches financial questions, load the skill.

730
00:25:04,520 --> 00:25:07,040
But that same description also matches questions about

731
00:25:07,040 --> 00:25:09,520
approving an expense or actual spend from last month

732
00:25:09,520 --> 00:25:11,160
or how to request a budget increase.

733
00:25:11,160 --> 00:25:12,760
All of these are financial.

734
00:25:12,760 --> 00:25:14,480
All of them treat are the same skill.

735
00:25:14,480 --> 00:25:17,200
But in reality, they are entirely different workflows.

736
00:25:17,200 --> 00:25:19,080
They have different tools, different constraints,

737
00:25:19,080 --> 00:25:20,080
and different outcomes.

738
00:25:20,080 --> 00:25:21,640
The skill loads when it shouldn't,

739
00:25:21,640 --> 00:25:23,360
or it loads and the model struggles

740
00:25:23,360 --> 00:25:25,280
because the instructions assume a specific type

741
00:25:25,280 --> 00:25:28,600
of financial question that doesn't match the user's intent.

742
00:25:28,600 --> 00:25:30,080
On the other extreme, a description

743
00:25:30,080 --> 00:25:31,720
that's too narrow never triggers.

744
00:25:31,720 --> 00:25:34,160
If you tell the agent to use a skill only for processing

745
00:25:34,160 --> 00:25:37,160
Q4 vendor reimbursements over $5,000 submitted

746
00:25:37,160 --> 00:25:39,200
after a specific budget freeze, it might

747
00:25:39,200 --> 00:25:41,200
match three scenarios per year.

748
00:25:41,200 --> 00:25:42,920
Users never get the benefit because they

749
00:25:42,920 --> 00:25:45,160
don't phrase their requests in those exact terms.

750
00:25:45,160 --> 00:25:48,000
The skill sits dormant, the routing fails silently.

751
00:25:48,000 --> 00:25:49,680
The distinction between routing metadata

752
00:25:49,680 --> 00:25:51,800
and documentation is critical here.

753
00:25:51,800 --> 00:25:54,320
Your skills instructions, the step-by-step procedures,

754
00:25:54,320 --> 00:25:55,600
those are documentation.

755
00:25:55,600 --> 00:25:57,360
They explain how to execute the workflow

756
00:25:57,360 --> 00:25:58,840
to a human reading the file.

757
00:25:58,840 --> 00:25:59,520
They can be long.

758
00:25:59,520 --> 00:26:01,360
They can include caveats and examples.

759
00:26:01,360 --> 00:26:03,000
The name and description are different.

760
00:26:03,000 --> 00:26:04,320
They are signals.

761
00:26:04,320 --> 00:26:06,080
They are the data the orchestrator uses

762
00:26:06,080 --> 00:26:07,600
to make a binary decision.

763
00:26:07,600 --> 00:26:09,160
Load this skill or don't.

764
00:26:09,160 --> 00:26:12,200
If the description is ambiguous, the orchestrator has to guess.

765
00:26:12,200 --> 00:26:13,720
If it's too narrow, it never matches.

766
00:26:13,720 --> 00:26:15,240
Precision here isn't optional.

767
00:26:15,240 --> 00:26:17,400
It determines whether the routing actually works.

768
00:26:17,400 --> 00:26:20,200
So how does the orchestrator actually select a skill?

769
00:26:20,200 --> 00:26:22,560
It reads the user's request and extracts the intent.

770
00:26:22,560 --> 00:26:24,920
It compares that intent against the name and description

771
00:26:24,920 --> 00:26:27,400
of every available skill and scores the matches.

772
00:26:27,400 --> 00:26:29,600
The skill with the highest relevance score loads.

773
00:26:29,600 --> 00:26:31,920
If the score is below a threshold, no skill loads,

774
00:26:31,920 --> 00:26:34,080
and the agent reasons from its baseline instructions.

775
00:26:34,080 --> 00:26:35,920
This is where the precision rule applies.

776
00:26:35,920 --> 00:26:37,640
If two reasonable makers would disagree

777
00:26:37,640 --> 00:26:40,320
about whether a skill should load for a particular request,

778
00:26:40,320 --> 00:26:42,400
the description isn't specific enough.

779
00:26:42,400 --> 00:26:45,200
You need to revise it until the boundary is unambiguous.

780
00:26:45,200 --> 00:26:47,200
The reasoning view is your main debugging surface

781
00:26:47,200 --> 00:26:48,240
for this problem.

782
00:26:48,240 --> 00:26:50,880
Inside Copilot Studio, you can open the reasoning view

783
00:26:50,880 --> 00:26:52,280
to see exactly what happened.

784
00:26:52,280 --> 00:26:54,920
You can see which skill loaded, why it was chosen,

785
00:26:54,920 --> 00:26:56,040
and what the matching score was.

786
00:26:56,040 --> 00:26:58,400
If the wrong skill loaded, the reasoning view shows you

787
00:26:58,400 --> 00:27:00,640
that the description didn't provide enough distinction

788
00:27:00,640 --> 00:27:02,880
or the name overlapped with another skill.

789
00:27:02,880 --> 00:27:04,760
You can watch the model's actual reasoning path.

790
00:27:04,760 --> 00:27:06,360
You see the choice and the justification.

791
00:27:06,360 --> 00:27:08,400
From there, you can refine the description,

792
00:27:08,400 --> 00:27:09,440
make it more specific.

793
00:27:09,440 --> 00:27:12,920
At exclusion language like, do not use for payroll questions.

794
00:27:12,920 --> 00:27:15,400
Clarify the trigger condition, narrow the scope.

795
00:27:15,400 --> 00:27:16,200
Itterate.

796
00:27:16,200 --> 00:27:18,480
Two broad causes constant unwanted firing.

797
00:27:18,480 --> 00:27:20,320
Two narrow causes routing failure.

798
00:27:20,320 --> 00:27:21,640
The balance point is precision.

799
00:27:21,640 --> 00:27:23,640
This is solvable and solved correctly.

800
00:27:23,640 --> 00:27:24,600
It's invisible.

801
00:27:24,600 --> 00:27:26,400
The skill loads exactly when it should.

802
00:27:26,400 --> 00:27:27,880
The user doesn't think about routing.

803
00:27:27,880 --> 00:27:30,280
They ask their question, the right expertise loads,

804
00:27:30,280 --> 00:27:32,280
and execution proceeds smoothly.

805
00:27:32,280 --> 00:27:33,560
But get the descriptions wrong

806
00:27:33,560 --> 00:27:35,880
and the entire architecture falls apart.

807
00:27:35,880 --> 00:27:38,520
When to use instructions versus skills?

808
00:27:38,520 --> 00:27:41,120
Here's the practical question every architect faces.

809
00:27:41,120 --> 00:27:42,680
Where does this guidance live?

810
00:27:42,680 --> 00:27:44,160
Does it go in the instructions?

811
00:27:44,160 --> 00:27:46,240
Or do I extract it into a skill?

812
00:27:46,240 --> 00:27:48,160
The answer hinges on a single question.

813
00:27:48,160 --> 00:27:50,040
Is this rule true in every conversation?

814
00:27:50,040 --> 00:27:52,000
If yes, it belongs in instructions.

815
00:27:52,000 --> 00:27:53,280
If no, it belongs in a skill.

816
00:27:53,280 --> 00:27:54,680
That's the decision tree.

817
00:27:54,680 --> 00:27:56,280
Instructions are the always on layer.

818
00:27:56,280 --> 00:27:57,600
They're loaded before any skill

819
00:27:57,600 --> 00:27:59,240
and before the user says anything.

820
00:27:59,240 --> 00:28:01,280
Instructions are what you want the agent to be,

821
00:28:01,280 --> 00:28:04,080
regardless of the scenario, your tone, your identity,

822
00:28:04,080 --> 00:28:05,320
the way you speak.

823
00:28:05,320 --> 00:28:08,440
You are helpful, direct, and professional is an instruction.

824
00:28:08,440 --> 00:28:10,560
It applies when handling leave requests

825
00:28:10,560 --> 00:28:13,080
and it applies when processing expense approvals.

826
00:28:13,080 --> 00:28:13,960
It's constant.

827
00:28:13,960 --> 00:28:15,720
It belongs in the instruction layer.

828
00:28:15,720 --> 00:28:17,360
Same with roll.

829
00:28:17,360 --> 00:28:19,680
You are an HR assistant sets your identity.

830
00:28:19,680 --> 00:28:21,440
That doesn't change between conversations.

831
00:28:21,440 --> 00:28:23,240
Universal guardrails belong here too.

832
00:28:23,240 --> 00:28:26,280
You might tell the agent to never promise legally binding outcomes

833
00:28:26,280 --> 00:28:29,040
or to always protect personal data according to policy.

834
00:28:29,040 --> 00:28:30,440
These rules apply universally.

835
00:28:30,440 --> 00:28:31,400
They're not situational.

836
00:28:31,400 --> 00:28:34,240
A leave eligibility skill doesn't need to restate them

837
00:28:34,240 --> 00:28:36,480
because the base instructions already hold them.

838
00:28:36,480 --> 00:28:37,840
The skill can assume they're true

839
00:28:37,840 --> 00:28:39,880
and focus on the specifics of leave.

840
00:28:39,880 --> 00:28:41,360
This is the manageability payoff.

841
00:28:41,360 --> 00:28:43,320
Your instructions stay lean because they're not trying

842
00:28:43,320 --> 00:28:45,760
to encode every procedure and every edge case.

843
00:28:45,760 --> 00:28:47,000
They're the common baseline.

844
00:28:47,000 --> 00:28:49,000
Everything situational lives elsewhere.

845
00:28:49,000 --> 00:28:52,240
Now flip the question, is this true only in specific scenarios?

846
00:28:52,240 --> 00:28:54,280
Situational guidance belongs in skills.

847
00:28:54,280 --> 00:28:56,480
The step-by-step procedure for processing a vendor refund

848
00:28:56,480 --> 00:28:58,240
above a certain threshold is situational.

849
00:28:58,240 --> 00:28:59,800
It applies to that specific scenario

850
00:28:59,800 --> 00:29:02,280
but it doesn't apply when processing employee reimbursements

851
00:29:02,280 --> 00:29:03,320
or customer refunds.

852
00:29:03,320 --> 00:29:05,880
It doesn't apply most of the time, so it lives in a skill.

853
00:29:05,880 --> 00:29:08,920
When a user asks about vendor refunds, the skill loads.

854
00:29:08,920 --> 00:29:10,200
When they ask about something else,

855
00:29:10,200 --> 00:29:11,680
the skill stays unloaded.

856
00:29:11,680 --> 00:29:13,320
The rules for approving a leave request

857
00:29:13,320 --> 00:29:15,520
during a blackout period are situational.

858
00:29:15,520 --> 00:29:16,720
They only matter during blackout

859
00:29:16,720 --> 00:29:18,120
and only for leave requests.

860
00:29:18,120 --> 00:29:18,800
Skill.

861
00:29:18,800 --> 00:29:20,240
The checklist for high-value contracts

862
00:29:20,240 --> 00:29:22,760
over $2 million with specific legal review requirements

863
00:29:22,760 --> 00:29:23,760
is situational.

864
00:29:23,760 --> 00:29:24,560
Skill.

865
00:29:24,560 --> 00:29:26,400
The difference is context dependency.

866
00:29:26,400 --> 00:29:28,080
Always true rules live in instructions.

867
00:29:28,080 --> 00:29:30,000
Context dependent rules live in skills,

868
00:29:30,000 --> 00:29:31,200
but here's the subtlety.

869
00:29:31,200 --> 00:29:33,880
Some guidance feels universal because it's complicated

870
00:29:33,880 --> 00:29:35,680
and you want the model to have it available.

871
00:29:35,680 --> 00:29:36,640
Resist that impulse.

872
00:29:36,640 --> 00:29:38,480
Complicated doesn't mean always on.

873
00:29:38,480 --> 00:29:41,720
Ask yourself if the user encounters this in every conversation.

874
00:29:41,720 --> 00:29:43,680
If the answer is no, it's situational.

875
00:29:43,680 --> 00:29:45,160
Extract it into a skill.

876
00:29:45,160 --> 00:29:46,600
Let's look at a concrete example.

877
00:29:46,600 --> 00:29:48,560
An organization builds an HR agent.

878
00:29:48,560 --> 00:29:51,080
The initial approach is to dump everything into the instructions,

879
00:29:51,080 --> 00:29:54,080
leave policy, salary questions, benefits enrollment,

880
00:29:54,080 --> 00:29:56,840
performance reviews, and compensation structure.

881
00:29:56,840 --> 00:29:59,120
They add payroll tax treatment, holiday schedules,

882
00:29:59,120 --> 00:30:01,160
remote work policies, and disability leave,

883
00:30:01,160 --> 00:30:03,440
all in one instruction file, thousands of lines.

884
00:30:03,440 --> 00:30:04,480
The instructions are so bloated

885
00:30:04,480 --> 00:30:06,960
that the model has to process all of it for every single turn.

886
00:30:06,960 --> 00:30:09,040
Most of it is irrelevant to most conversations.

887
00:30:09,040 --> 00:30:12,280
The alternative approach is to keep instructions minimal.

888
00:30:12,280 --> 00:30:14,600
Identity, you are the HR assistant.

889
00:30:14,600 --> 00:30:16,920
Tone, professional, empathetic, clear.

890
00:30:16,920 --> 00:30:18,280
Universal guardrails.

891
00:30:18,280 --> 00:30:20,320
All salary information is confidential.

892
00:30:20,320 --> 00:30:23,480
Never promise benefits decisions without HR review.

893
00:30:23,480 --> 00:30:24,600
That's the instruction layer.

894
00:30:24,600 --> 00:30:26,480
It's a hundred lines maximum.

895
00:30:26,480 --> 00:30:28,120
Now create three skills.

896
00:30:28,120 --> 00:30:30,840
One for leave eligibility, covering a cruel rules

897
00:30:30,840 --> 00:30:32,080
and blackout periods.

898
00:30:32,080 --> 00:30:35,320
One for compensation, covering salary grades and review cycles.

899
00:30:35,320 --> 00:30:38,240
One for enrollment, covering benefits and retirement plans.

900
00:30:38,240 --> 00:30:41,760
When someone asks about their PTO balance, the leave skill loads.

901
00:30:41,760 --> 00:30:43,760
The model has the specific rules for PTO,

902
00:30:43,760 --> 00:30:45,760
a cruel and company specific policies.

903
00:30:45,760 --> 00:30:47,240
So it answers with precision.

904
00:30:47,240 --> 00:30:49,960
When someone asks a different question, a different skill loads.

905
00:30:49,960 --> 00:30:51,320
The instructions are unchanged.

906
00:30:51,320 --> 00:30:53,640
The model's identity and values are consistent.

907
00:30:53,640 --> 00:30:55,360
But the procedural guidance is modular.

908
00:30:55,360 --> 00:30:57,120
The leave team owns the leave skill

909
00:30:57,120 --> 00:30:59,880
and the compensation team owns the compensation skill.

910
00:30:59,880 --> 00:31:02,560
Each can update their domain without touching the other.

911
00:31:02,560 --> 00:31:04,720
If leave policies change, you update one skill,

912
00:31:04,720 --> 00:31:06,280
not a massive instruction file.

913
00:31:06,280 --> 00:31:08,240
This distinction prevents context saturation.

914
00:31:08,240 --> 00:31:09,400
It keeps instructions lean.

915
00:31:09,400 --> 00:31:10,480
It distributes ownership.

916
00:31:10,480 --> 00:31:12,280
It makes the system maintainable at scale.

917
00:31:12,280 --> 00:31:13,480
The decision is simple.

918
00:31:13,480 --> 00:31:17,040
Universal, instructions, situational, skills.

919
00:31:17,040 --> 00:31:18,520
Everything follows from that.

920
00:31:18,520 --> 00:31:20,440
The skill versus new agent decision.

921
00:31:20,440 --> 00:31:22,840
So when does the skill actually become a separate agent?

922
00:31:22,840 --> 00:31:24,160
This is the inverse question.

923
00:31:24,160 --> 00:31:26,680
We've established when to pull guidance out of instructions

924
00:31:26,680 --> 00:31:27,840
and into a skill.

925
00:31:27,840 --> 00:31:28,840
Now comes the harder part.

926
00:31:28,840 --> 00:31:29,760
You've built a skill.

927
00:31:29,760 --> 00:31:30,400
It's working.

928
00:31:30,400 --> 00:31:32,280
Teams are using it, but now you're wondering

929
00:31:32,280 --> 00:31:33,920
if it's grown so large that it should just

930
00:31:33,920 --> 00:31:35,520
be its own standalone agent.

931
00:31:35,520 --> 00:31:36,840
The answer isn't about size.

932
00:31:36,840 --> 00:31:38,240
A skill can be comprehensive.

933
00:31:38,240 --> 00:31:39,880
It can have hundreds of lines of logic.

934
00:31:39,880 --> 00:31:42,440
It can bundle scripts, templates, and reference materials

935
00:31:42,440 --> 00:31:43,520
into one package.

936
00:31:43,520 --> 00:31:45,720
Size alone doesn't determine the boundary.

937
00:31:45,720 --> 00:31:48,760
The real distinction comes down to audience and knowledge domain.

938
00:31:48,760 --> 00:31:50,520
Think of it as an audience test.

939
00:31:50,520 --> 00:31:53,240
Does this skill serve the same users as the parent agent?

940
00:31:53,240 --> 00:31:54,800
Does it share the same knowledge boundary?

941
00:31:54,800 --> 00:31:57,240
When you have an HR agent, a leave eligibility skill

942
00:31:57,240 --> 00:31:59,840
serves HR users asking HR questions.

943
00:31:59,840 --> 00:32:01,360
The knowledge domain is consistent.

944
00:32:01,360 --> 00:32:03,240
The same people interact with leave questions,

945
00:32:03,240 --> 00:32:05,400
compensation questions, and benefits questions.

946
00:32:05,400 --> 00:32:06,360
It's all HR.

947
00:32:06,360 --> 00:32:07,960
A single agent with multiple skills

948
00:32:07,960 --> 00:32:10,560
makes sense here because the user population is unified

949
00:32:10,560 --> 00:32:12,120
and the domain is coherent.

950
00:32:12,120 --> 00:32:13,640
But if you find yourself building a skill

951
00:32:13,640 --> 00:32:15,160
that serves a completely different group,

952
00:32:15,160 --> 00:32:16,000
that's a signal.

953
00:32:16,000 --> 00:32:18,480
Or if it serves the same users, but for an entirely different

954
00:32:18,480 --> 00:32:20,200
business domain, that's also a signal.

955
00:32:20,200 --> 00:32:22,920
An IT agent handles password resets, access provisioning,

956
00:32:22,920 --> 00:32:24,000
and software requests.

957
00:32:24,000 --> 00:32:25,560
These are all IT operations.

958
00:32:25,560 --> 00:32:26,640
They serve IT users.

959
00:32:26,640 --> 00:32:28,200
One agent with multiple skills works

960
00:32:28,200 --> 00:32:30,240
because the audience and the domain align.

961
00:32:30,240 --> 00:32:31,760
Now consider a different scenario.

962
00:32:31,760 --> 00:32:32,840
You have an IT agent.

963
00:32:32,840 --> 00:32:34,760
You're building a skill for data governance.

964
00:32:34,760 --> 00:32:38,000
This covers policies about PII handling, data classification,

965
00:32:38,000 --> 00:32:39,120
and retention rules.

966
00:32:39,120 --> 00:32:40,920
That skill serves IT staff sure.

967
00:32:40,920 --> 00:32:43,440
But it also serves business analysts, project managers,

968
00:32:43,440 --> 00:32:44,600
and compliance teams.

969
00:32:44,600 --> 00:32:47,040
The knowledge domain isn't IT operations anymore.

970
00:32:47,040 --> 00:32:48,160
It's data governance.

971
00:32:48,160 --> 00:32:51,080
The user population extends far beyond the IT desk.

972
00:32:51,080 --> 00:32:53,400
The semantic boundary is fundamentally different.

973
00:32:53,400 --> 00:32:55,800
At that point, a separate data governance agent

974
00:32:55,800 --> 00:32:56,880
makes more sense.

975
00:32:56,880 --> 00:32:58,600
Different users, different expertise,

976
00:32:58,600 --> 00:33:00,960
different knowledge boundary.

977
00:33:00,960 --> 00:33:03,080
The toolset test adds another dimension.

978
00:33:03,080 --> 00:33:05,080
Research shows that accuracy can degrade

979
00:33:05,080 --> 00:33:07,280
as an agent's available toolset grows.

980
00:33:07,280 --> 00:33:09,880
There is a practical limit to what one agent can juggle.

981
00:33:09,880 --> 00:33:11,880
When you exceed that limit, you're better off splitting

982
00:33:11,880 --> 00:33:13,840
into separate agents, rather than piling

983
00:33:13,840 --> 00:33:15,040
more tools into one.

984
00:33:15,040 --> 00:33:17,880
An IT agent that resets passwords, orders hardware,

985
00:33:17,880 --> 00:33:20,000
and manages licenses has broad reach.

986
00:33:20,000 --> 00:33:22,800
But it's still unified by the toolset and the domain.

987
00:33:22,800 --> 00:33:25,600
Now add data governance tools, compliance reporting,

988
00:33:25,600 --> 00:33:26,560
and audit logging.

989
00:33:26,560 --> 00:33:27,640
You've crossed the threshold.

990
00:33:27,640 --> 00:33:29,560
The toolset has become so expansive

991
00:33:29,560 --> 00:33:31,880
that the agent's reliability starts to suffer.

992
00:33:31,880 --> 00:33:33,760
You're better off separating them into two agents

993
00:33:33,760 --> 00:33:35,840
with focused curated tool sets.

994
00:33:35,840 --> 00:33:37,360
Look at a concrete example.

995
00:33:37,360 --> 00:33:40,560
An organization's IT support handles password resets,

996
00:33:40,560 --> 00:33:42,920
access approvals, and incident triage.

997
00:33:42,920 --> 00:33:45,360
And these are three distinct workflows, three skills,

998
00:33:45,360 --> 00:33:46,240
one agent.

999
00:33:46,240 --> 00:33:47,760
The agent can root to the right skill

1000
00:33:47,760 --> 00:33:49,280
based on what the user wants.

1001
00:33:49,280 --> 00:33:51,040
When a user says they forgot their password,

1002
00:33:51,040 --> 00:33:52,480
the reset skill loads.

1003
00:33:52,480 --> 00:33:54,800
When they ask for access to the financial system,

1004
00:33:54,800 --> 00:33:56,200
the approval skill loads.

1005
00:33:56,200 --> 00:33:57,600
When they report a server outage,

1006
00:33:57,600 --> 00:33:58,920
the triage skill loads.

1007
00:33:58,920 --> 00:34:00,120
Three scenarios.

1008
00:34:00,120 --> 00:34:02,760
One agent, one unified user experience.

1009
00:34:02,760 --> 00:34:04,560
Contrast that with HR and IT support.

1010
00:34:04,560 --> 00:34:06,320
Yes, they're both internal operations,

1011
00:34:06,320 --> 00:34:09,040
but they serve different users in different business domains.

1012
00:34:09,040 --> 00:34:11,880
An HR agent and an IT agent should be separate.

1013
00:34:11,880 --> 00:34:13,880
Users don't switch between them randomly.

1014
00:34:13,880 --> 00:34:16,480
A user needing leave approval goes to the HR agent.

1015
00:34:16,480 --> 00:34:19,240
A user needing a password reset goes to the IT agent,

1016
00:34:19,240 --> 00:34:21,560
different entry points, different expertise.

1017
00:34:21,560 --> 00:34:24,200
The modularity principle is worth restating here.

1018
00:34:24,200 --> 00:34:27,320
Skills are about teaching an agent new expertise within a domain.

1019
00:34:27,320 --> 00:34:29,600
A skill adds capability without changing

1020
00:34:29,600 --> 00:34:31,800
the agent's fundamental identity.

1021
00:34:31,800 --> 00:34:34,040
A separate agent is what you build when the domain,

1022
00:34:34,040 --> 00:34:36,360
audience, or toolset becomes so large

1023
00:34:36,360 --> 00:34:38,520
that unification no longer makes sense.

1024
00:34:38,520 --> 00:34:40,600
When in doubt ask, would a user see this

1025
00:34:40,600 --> 00:34:42,080
as one cohesive service?

1026
00:34:42,080 --> 00:34:43,560
Would they naturally go to this one agent

1027
00:34:43,560 --> 00:34:44,920
for everything in this domain?

1028
00:34:44,920 --> 00:34:46,560
If the answer is yes, use skills.

1029
00:34:46,560 --> 00:34:49,280
If the answer is no, build separate agents.

1030
00:34:49,280 --> 00:34:51,680
Accuracy through targeted guidance.

1031
00:34:51,680 --> 00:34:55,760
Accuracy is where the real payoff of skills reveals itself.

1032
00:34:55,760 --> 00:34:57,800
And it's not obvious at first glance.

1033
00:34:57,800 --> 00:35:00,080
You might think accuracy depends entirely on the model.

1034
00:35:00,080 --> 00:35:02,280
Bigger models are smarter, better training data

1035
00:35:02,280 --> 00:35:05,280
produces better answers, fine tuning improves performance

1036
00:35:05,280 --> 00:35:06,640
on your specific domain.

1037
00:35:06,640 --> 00:35:09,160
Those things are true, but they aren't the whole story.

1038
00:35:09,160 --> 00:35:11,840
Model quality matters, but so does signal clarity.

1039
00:35:11,840 --> 00:35:14,920
This is where skills become a lever you can actually pull.

1040
00:35:14,920 --> 00:35:17,560
The problem with broad generic guidance is noise.

1041
00:35:17,560 --> 00:35:19,200
When you give an agent instructions covering

1042
00:35:19,200 --> 00:35:22,480
every possible scenario, you're bearing the critical details.

1043
00:35:22,480 --> 00:35:24,320
The model reads 100 lines of guidance

1044
00:35:24,320 --> 00:35:26,760
when it only needs three to do the job well.

1045
00:35:26,760 --> 00:35:29,240
The rest is just context it has to pass and dismiss.

1046
00:35:29,240 --> 00:35:30,800
Cognitively, that's expensive.

1047
00:35:30,800 --> 00:35:33,240
The model attends to relevant and irrelevant information

1048
00:35:33,240 --> 00:35:34,000
at the same time.

1049
00:35:34,000 --> 00:35:35,520
It has to decide what matters.

1050
00:35:35,520 --> 00:35:36,720
Sometimes it guesses wrong.

1051
00:35:36,720 --> 00:35:39,080
Consider a tool use scenario.

1052
00:35:39,080 --> 00:35:40,560
You have an approval workflow.

1053
00:35:40,560 --> 00:35:42,240
The model is about to trigger it.

1054
00:35:42,240 --> 00:35:44,040
There are parameters like threshold amounts,

1055
00:35:44,040 --> 00:35:46,720
required approval levels and notification lists.

1056
00:35:46,720 --> 00:35:48,720
There's an escalation path for urgent cases

1057
00:35:48,720 --> 00:35:51,240
and fallback behavior if people are unavailable.

1058
00:35:51,240 --> 00:35:53,640
Now, if all of that lives in a generic instruction set

1059
00:35:53,640 --> 00:35:56,320
alongside 10 other workflows, the model is processing

1060
00:35:56,320 --> 00:35:58,840
a lot of material just to finish one task.

1061
00:35:58,840 --> 00:36:01,240
It's reading about budget codes when it should be focused

1062
00:36:01,240 --> 00:36:04,120
on whether this specific request triggers an approval at all.

1063
00:36:04,120 --> 00:36:05,880
The signal to noise ratio is poor.

1064
00:36:05,880 --> 00:36:07,640
The model might forget a validation step.

1065
00:36:07,640 --> 00:36:09,120
It might apply the wrong threshold.

1066
00:36:09,120 --> 00:36:10,880
It might miss an escalation condition.

1067
00:36:10,880 --> 00:36:12,600
Contrast that with a targeted skill.

1068
00:36:12,600 --> 00:36:15,400
The skill exists specifically for approval workflows.

1069
00:36:15,400 --> 00:36:17,800
When it loads, every single line of instruction

1070
00:36:17,800 --> 00:36:18,920
is about approvals.

1071
00:36:18,920 --> 00:36:19,960
Nothing else.

1072
00:36:19,960 --> 00:36:21,320
The rules are explicit.

1073
00:36:21,320 --> 00:36:24,200
Amounts under 5,000 require level 1 approval.

1074
00:36:24,200 --> 00:36:26,840
Amounts over 50,000 require executive leadership.

1075
00:36:26,840 --> 00:36:28,160
The validation steps are clear.

1076
00:36:28,160 --> 00:36:29,120
Check the fields.

1077
00:36:29,120 --> 00:36:30,280
Verify the requester.

1078
00:36:30,280 --> 00:36:31,920
Confirm the level exists.

1079
00:36:31,920 --> 00:36:33,600
The model process is targeted content.

1080
00:36:33,600 --> 00:36:35,080
High signal, no noise.

1081
00:36:35,080 --> 00:36:37,000
Every line moves the task forward.

1082
00:36:37,000 --> 00:36:38,880
The accuracy of the tool invocation improves

1083
00:36:38,880 --> 00:36:40,560
because the guidance is concentrated.

1084
00:36:40,560 --> 00:36:43,680
There's no cognitive load from passing irrelevant material.

1085
00:36:43,680 --> 00:36:45,560
There's no ambiguity about which rules apply.

1086
00:36:45,560 --> 00:36:47,080
The model executes with precision.

1087
00:36:47,080 --> 00:36:48,360
This isn't just theory.

1088
00:36:48,360 --> 00:36:50,480
Teams that implement skills report higher accuracy

1089
00:36:50,480 --> 00:36:53,720
on specific tasks compared to relying on broad documentation.

1090
00:36:53,720 --> 00:36:55,320
The improvement isn't universal.

1091
00:36:55,320 --> 00:36:57,440
It depends on the task, the tool set,

1092
00:36:57,440 --> 00:36:59,080
and how well the skill is written.

1093
00:36:59,080 --> 00:37:01,080
But narrow, curated skills consistently

1094
00:37:01,080 --> 00:37:03,520
outperform broad guidance for specialized workflows.

1095
00:37:03,520 --> 00:37:05,320
The distinction matters between accuracy

1096
00:37:05,320 --> 00:37:08,080
gains that are structural and those that are use case dependent.

1097
00:37:08,080 --> 00:37:09,600
Some improvements are structural.

1098
00:37:09,600 --> 00:37:11,440
When you remove noise and leave signal,

1099
00:37:11,440 --> 00:37:13,400
the model's reasoning path is clearer.

1100
00:37:13,400 --> 00:37:14,680
That's universal.

1101
00:37:14,680 --> 00:37:16,800
But other improvements depend on whether your skill

1102
00:37:16,800 --> 00:37:19,240
actually captures something the model struggles with.

1103
00:37:19,240 --> 00:37:20,880
If the model would get the task right,

1104
00:37:20,880 --> 00:37:23,520
even with generic guidance, a skill won't help much.

1105
00:37:23,520 --> 00:37:24,840
It's already accurate.

1106
00:37:24,840 --> 00:37:28,040
The skill adds clarity, but not necessarily a performance boost.

1107
00:37:28,040 --> 00:37:30,600
You have to measure to know which category you're in.

1108
00:37:30,600 --> 00:37:33,280
This is where the evaluation requirement becomes critical.

1109
00:37:33,280 --> 00:37:35,240
Before you move a workflow into a skill,

1110
00:37:35,240 --> 00:37:36,400
measure the baseline.

1111
00:37:36,400 --> 00:37:39,160
How often does the model get it right with generic guidance?

1112
00:37:39,160 --> 00:37:41,720
How often does it make mistakes document the failure modes?

1113
00:37:41,720 --> 00:37:44,120
Then implement the skill, run the same test cases

1114
00:37:44,120 --> 00:37:45,800
through the skill-driven workflow.

1115
00:37:45,800 --> 00:37:47,280
Measure again.

1116
00:37:47,280 --> 00:37:49,040
Did accuracy improve by how much?

1117
00:37:49,040 --> 00:37:51,040
Is the improvement consistent across different types

1118
00:37:51,040 --> 00:37:53,320
of requests or only in specific cases?

1119
00:37:53,320 --> 00:37:54,760
Don't assume skills help.

1120
00:37:54,760 --> 00:37:56,120
Measure.

1121
00:37:56,120 --> 00:37:58,760
Some tasks are already handled well by a capable model

1122
00:37:58,760 --> 00:38:00,040
with basic guidance.

1123
00:38:00,040 --> 00:38:01,440
Adding a skill won't move the needle

1124
00:38:01,440 --> 00:38:03,040
because the baseline is already high.

1125
00:38:03,040 --> 00:38:03,920
That's fine.

1126
00:38:03,920 --> 00:38:06,080
You've still gained from modularity and manageability

1127
00:38:06,080 --> 00:38:07,920
even if accuracy stayed the same.

1128
00:38:07,920 --> 00:38:09,680
Other tasks are fundamentally difficult.

1129
00:38:09,680 --> 00:38:11,880
A skill helps, but it doesn't solve everything.

1130
00:38:11,880 --> 00:38:14,080
The model still makes mistakes just fewer of them.

1131
00:38:14,080 --> 00:38:14,880
That's still a win.

1132
00:38:14,880 --> 00:38:16,440
And some tasks improve dramatically

1133
00:38:16,440 --> 00:38:18,040
because the skill removes confusion

1134
00:38:18,040 --> 00:38:19,840
and clarifies the exact steps.

1135
00:38:19,840 --> 00:38:21,800
The measurement discipline keeps you honest.

1136
00:38:21,800 --> 00:38:24,720
It prevents the assumption that skills are a silver bullet.

1137
00:38:24,720 --> 00:38:25,760
There are two.

1138
00:38:25,760 --> 00:38:28,280
They work when the problem is signal to noise clarity.

1139
00:38:28,280 --> 00:38:31,120
They work less well when the problem is an inadequate model

1140
00:38:31,120 --> 00:38:33,400
or a task that's inherently ambiguous.

1141
00:38:33,400 --> 00:38:34,440
Know the difference.

1142
00:38:34,440 --> 00:38:35,840
Measure before and after.

1143
00:38:35,840 --> 00:38:38,480
Let the data guide where you invest in skill development.

1144
00:38:38,480 --> 00:38:40,120
Speed and cost economics.

1145
00:38:40,120 --> 00:38:41,520
Here is the business reality.

1146
00:38:41,520 --> 00:38:43,840
Most engineers do not lead with.

1147
00:38:43,840 --> 00:38:45,560
Running agents at scale is expensive.

1148
00:38:45,560 --> 00:38:47,560
Every inference costs tokens.

1149
00:38:47,560 --> 00:38:49,320
Every token costs money.

1150
00:38:49,320 --> 00:38:50,760
In the world of co-pilot credits,

1151
00:38:50,760 --> 00:38:53,600
every single conversation is a line item on your budget.

1152
00:38:53,600 --> 00:38:56,240
When you are running hundreds of conversations every day,

1153
00:38:56,240 --> 00:38:58,160
that arithmetic starts to matter.

1154
00:38:58,160 --> 00:39:01,120
A single slow conversation that takes twice as long to finish

1155
00:39:01,120 --> 00:39:02,880
does more than just annoy your users.

1156
00:39:02,880 --> 00:39:05,640
It doubles the token consumption for that one interaction.

1157
00:39:05,640 --> 00:39:08,280
If you multiply that across your entire user base,

1158
00:39:08,280 --> 00:39:11,000
you are looking at a massive spike in operational spend.

1159
00:39:11,000 --> 00:39:12,680
Skills change the economics because they

1160
00:39:12,680 --> 00:39:14,280
change the token equation.

1161
00:39:14,280 --> 00:39:15,560
Think about the per turn math.

1162
00:39:15,560 --> 00:39:17,240
In a monolithic architecture,

1163
00:39:17,240 --> 00:39:21,360
every single response turn has to process the entire instruction set.

1164
00:39:21,360 --> 00:39:24,400
Let's say your baseline instruction file is 5,000 tokens.

1165
00:39:24,400 --> 00:39:26,200
No matter what the user actually asks for,

1166
00:39:26,200 --> 00:39:28,920
those 5,000 tokens are processed every time.

1167
00:39:28,920 --> 00:39:30,840
The model reads them, the API counts them.

1168
00:39:30,840 --> 00:39:31,800
They show up on your bill.

1169
00:39:31,800 --> 00:39:32,880
But here is the shift.

1170
00:39:32,880 --> 00:39:34,280
When you implement skills,

1171
00:39:34,280 --> 00:39:36,280
you can reduce those baseline instructions

1172
00:39:36,280 --> 00:39:37,880
to maybe 1,000 tokens.

1173
00:39:37,880 --> 00:39:40,720
You keep the core identity and the universal guardrails there,

1174
00:39:40,720 --> 00:39:42,200
but everything else moves.

1175
00:39:42,200 --> 00:39:44,320
The specific guidance lives in the skills.

1176
00:39:44,320 --> 00:39:46,400
On average, when a user asks a question,

1177
00:39:46,400 --> 00:39:48,400
only one relevant skill loads.

1178
00:39:48,400 --> 00:39:51,480
If that skill is 800 tokens, your total context for that turn

1179
00:39:51,480 --> 00:39:53,040
is now 1,900 tokens.

1180
00:39:53,040 --> 00:39:56,040
You just reduce your per turn token count by more than half,

1181
00:39:56,040 --> 00:39:57,600
the math compounds quickly.

1182
00:39:57,600 --> 00:40:00,160
If your agent handles 1,000 conversations a day

1183
00:40:00,160 --> 00:40:03,520
with 10 turns each, you are looking at 10,000 turns daily.

1184
00:40:03,520 --> 00:40:04,920
At 5,000 tokens per turn,

1185
00:40:04,920 --> 00:40:07,120
you are burning 50 million tokens every day.

1186
00:40:07,120 --> 00:40:09,080
But if you reduce that to 1,900 tokens,

1187
00:40:09,080 --> 00:40:10,680
you drop to 1,900,000.

1188
00:40:10,680 --> 00:40:12,000
That difference is not marginal.

1189
00:40:12,000 --> 00:40:13,120
It is structural.

1190
00:40:13,120 --> 00:40:15,760
Lower token counts lead directly to lower latency.

1191
00:40:15,760 --> 00:40:18,800
The model process is faster because it has less to read.

1192
00:40:18,800 --> 00:40:20,840
This also helps you scale your throughput.

1193
00:40:20,840 --> 00:40:22,400
With the same computational budget,

1194
00:40:22,400 --> 00:40:23,880
you can handle more people at once,

1195
00:40:23,880 --> 00:40:27,720
because each conversation is not eating through tokens so aggressively.

1196
00:40:27,720 --> 00:40:29,040
But here is where things change.

1197
00:40:29,040 --> 00:40:30,640
The throughput under load is where skills

1198
00:40:30,640 --> 00:40:32,120
give you the most leverage.

1199
00:40:32,120 --> 00:40:34,560
Most agents do not slow down gradually.

1200
00:40:34,560 --> 00:40:37,320
They fail catastrophically once they hit a limit.

1201
00:40:37,320 --> 00:40:40,000
You might be running fine with 50 users, but at 60,

1202
00:40:40,000 --> 00:40:41,240
you hit a quota wall.

1203
00:40:41,240 --> 00:40:42,440
Request start to queue.

1204
00:40:42,440 --> 00:40:43,960
Response times spike.

1205
00:40:43,960 --> 00:40:45,640
The whole system feels broken.

1206
00:40:45,640 --> 00:40:48,640
Skills prevent this by keeping the baseline context window lean.

1207
00:40:48,640 --> 00:40:52,280
At 50 users, you are not consuming 10 times the tokens

1208
00:40:52,280 --> 00:40:53,800
you would use for five users.

1209
00:40:53,800 --> 00:40:55,440
You are consuming proportionally less

1210
00:40:55,440 --> 00:40:57,720
because the context changes based on the task.

1211
00:40:57,720 --> 00:40:59,760
When the load spikes, you have more room to breathe

1212
00:40:59,760 --> 00:41:01,800
before you hit your infrastructure limits.

1213
00:41:01,800 --> 00:41:03,960
This changes how you plan for capacity.

1214
00:41:03,960 --> 00:41:06,800
Instead of asking how many users a configuration can support,

1215
00:41:06,800 --> 00:41:08,920
you ask what the per user token budget is.

1216
00:41:08,920 --> 00:41:11,080
Because skills let you right size the allocation

1217
00:41:11,080 --> 00:41:13,360
for every conversation, you can pack more users

1218
00:41:13,360 --> 00:41:14,760
into your existing capacity.

1219
00:41:14,760 --> 00:41:17,120
The impact on co-pilot credits is easy to see.

1220
00:41:17,120 --> 00:41:19,360
Fewer tokens per turn and fewer turns per task

1221
00:41:19,360 --> 00:41:21,800
mean you burn fewer credits to reach the goal.

1222
00:41:21,800 --> 00:41:24,800
Organizations that move away from bloated instruction files

1223
00:41:24,800 --> 00:41:28,000
often see a 60% reduction in tokens loaded per turn.

1224
00:41:28,000 --> 00:41:29,520
If your costs are driven by tokens,

1225
00:41:29,520 --> 00:41:32,400
that is a 60% direct saving on your workload.

1226
00:41:32,400 --> 00:41:34,560
The always on everything model is an anti-pattern

1227
00:41:34,560 --> 00:41:36,120
that destroys these economics.

1228
00:41:36,120 --> 00:41:39,080
When every agent loads every possible piece of guidance

1229
00:41:39,080 --> 00:41:41,120
permanently, you are paying for knowledge

1230
00:41:41,120 --> 00:41:42,440
the user will never need.

1231
00:41:42,440 --> 00:41:44,120
You are paying for a full reference manual

1232
00:41:44,120 --> 00:41:45,960
when the user just needs a checklist.

1233
00:41:45,960 --> 00:41:48,080
You are paying to load specialized procedures

1234
00:41:48,080 --> 00:41:49,800
that do not apply to the situation,

1235
00:41:49,800 --> 00:41:52,320
that overhead adds up across every single conversation

1236
00:41:52,320 --> 00:41:53,440
in your enterprise.

1237
00:41:53,440 --> 00:41:55,320
The real-world benchmarks prove this works.

1238
00:41:55,320 --> 00:41:58,360
Organizations moving to modular skills report lower costs

1239
00:41:58,360 --> 00:42:01,200
per user and per task across their entire deployment.

1240
00:42:01,200 --> 00:42:03,520
When you combine lower consumption with faster speeds

1241
00:42:03,520 --> 00:42:06,160
and higher capacity, the economic argument for skills

1242
00:42:06,160 --> 00:42:07,360
is impossible to ignore.

1243
00:42:07,360 --> 00:42:09,240
It is not about optimization theater.

1244
00:42:09,240 --> 00:42:10,960
It is about fundamental efficiency.

1245
00:42:10,960 --> 00:42:12,720
Load what is needed, pay for what is loaded,

1246
00:42:12,720 --> 00:42:14,720
scale by keeping the default lean.

1247
00:42:14,720 --> 00:42:17,280
The skill bundle, scripts and resources.

1248
00:42:17,280 --> 00:42:19,120
A skill is not just a markdown file.

1249
00:42:19,120 --> 00:42:20,000
It is a package.

1250
00:42:20,000 --> 00:42:21,520
Understanding what belongs in that package

1251
00:42:21,520 --> 00:42:22,920
is the difference between a demo

1252
00:42:22,920 --> 00:42:24,440
and a production-ready tool.

1253
00:42:24,440 --> 00:42:26,240
It starts with the skill MD file.

1254
00:42:26,240 --> 00:42:27,160
This is your manifest.

1255
00:42:27,160 --> 00:42:30,040
It holds the metadata like the name, version and author

1256
00:42:30,040 --> 00:42:31,840
that the orchestrator uses for rooting.

1257
00:42:31,840 --> 00:42:33,320
Below that is the instruction section.

1258
00:42:33,320 --> 00:42:34,440
This is narrative content.

1259
00:42:34,440 --> 00:42:36,800
It is the procedural guidance and the decision trees

1260
00:42:36,800 --> 00:42:39,800
that the model reads to understand how to do the job.

1261
00:42:39,800 --> 00:42:40,800
But here is the rule.

1262
00:42:40,800 --> 00:42:42,280
Code does not go in the skill.

1263
00:42:42,280 --> 00:42:43,400
MD file.

1264
00:42:43,400 --> 00:42:45,680
Raw scripts, binary data and complex templates

1265
00:42:45,680 --> 00:42:46,760
live separately.

1266
00:42:46,760 --> 00:42:48,640
They are bundled alongside the manifest.

1267
00:42:48,640 --> 00:42:51,120
Scripts are the executable code, the skill calls

1268
00:42:51,120 --> 00:42:52,160
while it is working.

1269
00:42:52,160 --> 00:42:53,920
This could be Python or shell scripts.

1270
00:42:53,920 --> 00:42:56,120
When a skill needs to validate a data structure

1271
00:42:56,120 --> 00:42:58,720
or generate a report, it calls a bundled script.

1272
00:42:58,720 --> 00:43:00,760
The script runs and returns an answer.

1273
00:43:00,760 --> 00:43:03,120
And then the model uses that result in its reasoning.

1274
00:43:03,120 --> 00:43:05,120
We do not put the script in the instructions

1275
00:43:05,120 --> 00:43:07,840
because the model is not supposed to read it as a story.

1276
00:43:07,840 --> 00:43:09,280
It is supposed to execute it.

1277
00:43:09,280 --> 00:43:10,800
Imagine a vendor onboarding skill.

1278
00:43:10,800 --> 00:43:12,720
The instructions explain the steps,

1279
00:43:12,720 --> 00:43:15,440
like collecting info and running background checks.

1280
00:43:15,440 --> 00:43:17,240
But one step requires checking a tax ID

1281
00:43:17,240 --> 00:43:18,720
against the government database.

1282
00:43:18,720 --> 00:43:20,600
That logic is too complex for a tax prompt,

1283
00:43:20,600 --> 00:43:22,560
so you bundle a Python script instead.

1284
00:43:22,560 --> 00:43:24,440
When the skill runs, it triggers the script.

1285
00:43:24,440 --> 00:43:26,040
The script handles the API call.

1286
00:43:26,040 --> 00:43:28,520
The result comes back and the instructions continue.

1287
00:43:28,520 --> 00:43:30,600
The script stays in the bundle because it needs to travel

1288
00:43:30,600 --> 00:43:31,440
with the skill.

1289
00:43:31,440 --> 00:43:34,280
When you move the skill to production, the script moves to.

1290
00:43:34,280 --> 00:43:36,560
When you version the skill, you version the script.

1291
00:43:36,560 --> 00:43:39,160
If you revoke access to the skill, the script is gone.

1292
00:43:39,160 --> 00:43:40,600
They are coupled by design.

1293
00:43:40,600 --> 00:43:42,680
References are the documents the agent reads

1294
00:43:42,680 --> 00:43:43,920
but does not execute.

1295
00:43:43,920 --> 00:43:46,440
This includes things like schemers, policy documents,

1296
00:43:46,440 --> 00:43:47,720
or reference tables.

1297
00:43:47,720 --> 00:43:50,280
A skill for customer data might bundle a schema file

1298
00:43:50,280 --> 00:43:53,240
that describes the fields and relationships in your database.

1299
00:43:53,240 --> 00:43:56,040
The model reads this to understand the data it is looking at.

1300
00:43:56,040 --> 00:43:58,000
It is reference material, not code.

1301
00:43:58,000 --> 00:43:59,920
Assets are your templates and sample data.

1302
00:43:59,920 --> 00:44:01,400
A contract review skill might include

1303
00:44:01,400 --> 00:44:03,520
a template of standard company terms.

1304
00:44:03,520 --> 00:44:06,480
An expense report skill might bundle a sample report

1305
00:44:06,480 --> 00:44:08,480
to show the model the expected format.

1306
00:44:08,480 --> 00:44:09,760
These assets load on demand.

1307
00:44:09,760 --> 00:44:12,120
They are part of the package, but they only appear

1308
00:44:12,120 --> 00:44:13,400
when they are called.

1309
00:44:13,400 --> 00:44:16,520
It is important to understand how this loading actually works.

1310
00:44:16,520 --> 00:44:18,800
When the orchestrator picks a skill, it loads the skill.

1311
00:44:18,800 --> 00:44:20,560
MD instructions into the context immediately,

1312
00:44:20,560 --> 00:44:22,640
the model starts thinking, but the scripts

1313
00:44:22,640 --> 00:44:24,640
and the reference files do not load yet.

1314
00:44:24,640 --> 00:44:26,960
They are available, but they do not consume tokens

1315
00:44:26,960 --> 00:44:28,800
until the agent actually needs them.

1316
00:44:28,800 --> 00:44:30,720
If a conversation finishes without ever needing

1317
00:44:30,720 --> 00:44:33,960
that tax validation script, that code was never tokenized.

1318
00:44:33,960 --> 00:44:35,320
It never cost you a cent.

1319
00:44:35,320 --> 00:44:37,160
That is the efficiency of the bundle.

1320
00:44:37,160 --> 00:44:38,920
The package travels as one unit,

1321
00:44:38,920 --> 00:44:40,920
but the parts load only when necessary.

1322
00:44:40,920 --> 00:44:42,960
We also have to talk about the security surface.

1323
00:44:42,960 --> 00:44:45,000
Bundled scripts are untrusted code.

1324
00:44:45,000 --> 00:44:46,840
When you pull a skill from an external registry,

1325
00:44:46,840 --> 00:44:49,400
those scripts could contain anything from malicious payloads

1326
00:44:49,400 --> 00:44:50,520
to credential theft.

1327
00:44:50,520 --> 00:44:51,760
You cannot just let them run.

1328
00:44:51,760 --> 00:44:54,200
You have to treat them like any third party code.

1329
00:44:54,200 --> 00:44:56,680
You review them statically, run them in a sandbox,

1330
00:44:56,680 --> 00:44:58,080
and watch how they behave.

1331
00:44:58,080 --> 00:45:00,720
If a script looks suspicious, you do not load it.

1332
00:45:00,720 --> 00:45:02,200
The skill can usually still function

1333
00:45:02,200 --> 00:45:04,200
because the instructions are independent,

1334
00:45:04,200 --> 00:45:05,920
but if you are going to run bundled code,

1335
00:45:05,920 --> 00:45:07,000
you need real governance.

1336
00:45:07,000 --> 00:45:10,040
This is why enterprise governance focuses on the whole bundle.

1337
00:45:10,040 --> 00:45:11,560
The instructions are usually safe,

1338
00:45:11,560 --> 00:45:14,640
but the scripts are where the real risk lives.

1339
00:45:14,640 --> 00:45:17,280
The governance crisis, supply chain risk.

1340
00:45:17,280 --> 00:45:19,200
Up to this point, we have talked about skills

1341
00:45:19,200 --> 00:45:21,840
as a technical fix for architectural problems.

1342
00:45:21,840 --> 00:45:23,320
Better performance, lower costs,

1343
00:45:23,320 --> 00:45:25,800
and easier management are all part of the story.

1344
00:45:25,800 --> 00:45:27,040
But there is a darker reality

1345
00:45:27,040 --> 00:45:29,280
that enterprise architects are just starting to face,

1346
00:45:29,280 --> 00:45:31,400
and it changes everything about how you put skills

1347
00:45:31,400 --> 00:45:32,320
into production.

1348
00:45:32,320 --> 00:45:33,160
Skills are code.

1349
00:45:33,160 --> 00:45:35,080
The instructions themselves are just guidance

1350
00:45:35,080 --> 00:45:37,240
or narrative content for a model to read,

1351
00:45:37,240 --> 00:45:39,320
but skills also bundle executable scripts

1352
00:45:39,320 --> 00:45:41,120
like Python and Shell commands.

1353
00:45:41,120 --> 00:45:43,640
Once you allow executable code into an agent framework,

1354
00:45:43,640 --> 00:45:45,440
you have opened a supply chain attack surface

1355
00:45:45,440 --> 00:45:47,720
that most organizations are not ready to defend.

1356
00:45:47,720 --> 00:45:49,160
The evidence is alarming.

1357
00:45:49,160 --> 00:45:51,400
A deep analysis of nearly 100,000 skills

1358
00:45:51,400 --> 00:45:55,440
from public registries found 157 confirmed malicious skills.

1359
00:45:55,440 --> 00:45:57,680
These were not just suspicious or poorly written,

1360
00:45:57,680 --> 00:45:59,960
but were specifically designed to steal credentials

1361
00:45:59,960 --> 00:46:01,360
and exfiltrate data.

1362
00:46:01,360 --> 00:46:04,400
While 157 attacks were verified in this first wave

1363
00:46:04,400 --> 00:46:07,160
of adoption, the actual number is likely much higher

1364
00:46:07,160 --> 00:46:10,280
because many skills enclosed registries were never analyzed.

1365
00:46:10,280 --> 00:46:11,960
These attack patterns follow the same playbooks

1366
00:46:11,960 --> 00:46:13,880
we see in software supply chains.

1367
00:46:13,880 --> 00:46:16,040
Attackers use typosquoting by publishing a skill

1368
00:46:16,040 --> 00:46:19,440
named vendor onboarding, and hoping users do not notice

1369
00:46:19,440 --> 00:46:21,880
the misspelling when they are looking for the real vendor

1370
00:46:21,880 --> 00:46:23,880
onboarding skill.

1371
00:46:23,880 --> 00:46:25,640
Other skills use credential harvesting

1372
00:46:25,640 --> 00:46:28,280
to log authentication tokens, while some focus

1373
00:46:28,280 --> 00:46:30,560
on privilege escalation to request higher permissions

1374
00:46:30,560 --> 00:46:31,960
than they actually need.

1375
00:46:31,960 --> 00:46:35,040
Some skills even look legitimate and perform their stated task,

1376
00:46:35,040 --> 00:46:36,960
but they secretly copy sensitive data

1377
00:46:36,960 --> 00:46:39,600
to external endpoints while they are running.

1378
00:46:39,600 --> 00:46:42,000
The industry is moving toward a tiered trust model

1379
00:46:42,000 --> 00:46:43,080
to handle this.

1380
00:46:43,080 --> 00:46:45,480
Green zone skills come from verified vendors

1381
00:46:45,480 --> 00:46:48,120
and are signed cryptographically after being reviewed

1382
00:46:48,120 --> 00:46:49,400
by a security team.

1383
00:46:49,400 --> 00:46:51,800
When you import a green skill, you are trusting

1384
00:46:51,800 --> 00:46:54,520
that the vendor did their job, which is a strong signal,

1385
00:46:54,520 --> 00:46:56,720
even if it is not a perfect guarantee.

1386
00:46:56,720 --> 00:46:59,040
Amber zone skills are contributed by the community

1387
00:46:59,040 --> 00:47:01,640
or known authors who have not been fully verified.

1388
00:47:01,640 --> 00:47:03,480
Someone in the community has reviewed them,

1389
00:47:03,480 --> 00:47:05,160
but that review might have been thorough

1390
00:47:05,160 --> 00:47:06,640
or it might have been cursory.

1391
00:47:06,640 --> 00:47:08,560
Many of these skills are perfectly fine,

1392
00:47:08,560 --> 00:47:10,240
but because the assurance is lower,

1393
00:47:10,240 --> 00:47:12,200
you're accepting more risk by using them.

1394
00:47:12,200 --> 00:47:14,120
Red zone skills are completely un-vetted

1395
00:47:14,120 --> 00:47:15,560
and often published anonymously.

1396
00:47:15,560 --> 00:47:17,960
There is no review and no way to know what they do

1397
00:47:17,960 --> 00:47:20,200
without reading every line of code yourself.

1398
00:47:20,200 --> 00:47:23,320
Red skills simply do not belong in production systems.

1399
00:47:23,320 --> 00:47:25,440
The problem is that 60% of enterprise agents

1400
00:47:25,440 --> 00:47:28,560
currently in use have arbitrary code execution enabled.

1401
00:47:28,560 --> 00:47:31,360
Most organizations have not set up tiered governance yet,

1402
00:47:31,360 --> 00:47:33,920
so they are importing skills without reviews

1403
00:47:33,920 --> 00:47:35,840
and running red zone code in production

1404
00:47:35,840 --> 00:47:37,200
without even knowing it.

1405
00:47:37,200 --> 00:47:39,840
This creates a forensic black hole when something goes wrong.

1406
00:47:39,840 --> 00:47:42,000
In traditional software, you have deployment records

1407
00:47:42,000 --> 00:47:44,320
and audit trails to see what happened during a breach,

1408
00:47:44,320 --> 00:47:47,240
but in many skill deployments, those records do not exist.

1409
00:47:47,240 --> 00:47:49,040
A skill loads and a script runs,

1410
00:47:49,040 --> 00:47:51,400
but there is no log of file operations

1411
00:47:51,400 --> 00:47:53,800
or network calls to show what actually happened.

1412
00:47:53,800 --> 00:47:56,280
If a script is malicious or a good script gets compromised,

1413
00:47:56,280 --> 00:47:59,440
you are flying blind with no way to investigate the damage.

1414
00:47:59,440 --> 00:48:01,960
The only real antidote is a strict approval workflow.

1415
00:48:01,960 --> 00:48:03,760
Before any skill reaches production,

1416
00:48:03,760 --> 00:48:05,440
it must pass through static analysis

1417
00:48:05,440 --> 00:48:08,640
to scan for hard coded credentials and unusual system calls.

1418
00:48:08,640 --> 00:48:10,880
It needs to run in a sandbox environment

1419
00:48:10,880 --> 00:48:13,040
where you can observe if it accesses systems,

1420
00:48:13,040 --> 00:48:15,000
it should not or leaks data.

1421
00:48:15,000 --> 00:48:16,560
Finally, you need runtime monitoring

1422
00:48:16,560 --> 00:48:18,520
to log every material action the skill takes

1423
00:48:18,520 --> 00:48:20,840
so you have a record if things go sideways.

1424
00:48:20,840 --> 00:48:23,440
This governance framework is not a nice to have feature.

1425
00:48:23,440 --> 00:48:26,120
Organizations that deploy skills without these rules

1426
00:48:26,120 --> 00:48:28,720
are just betting that attackers have not found them yet,

1427
00:48:28,720 --> 00:48:31,360
and that is a vulnerability, not a strategy.

1428
00:48:31,360 --> 00:48:33,920
Zoned governance architecture,

1429
00:48:33,920 --> 00:48:35,640
knowing that skills need governance is one thing

1430
00:48:35,640 --> 00:48:38,320
but building the framework to enforce it is another.

1431
00:48:38,320 --> 00:48:41,640
The pattern that actually works at scale is environmental zoning.

1432
00:48:41,640 --> 00:48:44,160
This is not a security add on you bolt on at the end,

1433
00:48:44,160 --> 00:48:45,600
but the foundational structure

1434
00:48:45,600 --> 00:48:48,200
for how skills flow through your entire organization.

1435
00:48:48,200 --> 00:48:50,440
Everything starts with your environment strategy.

1436
00:48:50,440 --> 00:48:52,120
You are likely already using development,

1437
00:48:52,120 --> 00:48:54,480
test and production environments in the power platform,

1438
00:48:54,480 --> 00:48:57,120
and that structure now becomes your governance backbone.

1439
00:48:57,120 --> 00:48:59,800
Each environment accesses a zone with its own rules

1440
00:48:59,800 --> 00:49:02,680
about which skills can run and who is allowed to publish them.

1441
00:49:02,680 --> 00:49:06,040
Green zones are your spaces for experimentation and innovation.

1442
00:49:06,040 --> 00:49:07,440
These are development environments

1443
00:49:07,440 --> 00:49:10,200
where makers build new skills and test out different workflows.

1444
00:49:10,200 --> 00:49:11,840
The default setting here is permissive

1445
00:49:11,840 --> 00:49:14,240
so teams can import skills from public registries

1446
00:49:14,240 --> 00:49:16,120
and test them against development data.

1447
00:49:16,120 --> 00:49:18,000
The goal is to let people try things safely,

1448
00:49:18,000 --> 00:49:20,080
but safety still requires some control.

1449
00:49:20,080 --> 00:49:22,800
You still apply DLP policies to block dangerous connectors

1450
00:49:22,800 --> 00:49:24,600
like unrestricted HTTP calls,

1451
00:49:24,600 --> 00:49:27,320
but you let teams experiment with internal systems

1452
00:49:27,320 --> 00:49:30,640
because any failure is isolated to that one environment.

1453
00:49:30,640 --> 00:49:33,160
Yellow zones are where multiple makers collaborate

1454
00:49:33,160 --> 00:49:35,640
on skills that are heading toward production.

1455
00:49:35,640 --> 00:49:36,840
Governance gets tighter here

1456
00:49:36,840 --> 00:49:38,720
and only curated skills that have been reviewed

1457
00:49:38,720 --> 00:49:40,760
by a competent person are allowed to enter.

1458
00:49:40,760 --> 00:49:42,840
This is not a full corporate approval yet,

1459
00:49:42,840 --> 00:49:45,480
but it is a checkpoint where an architect confirms

1460
00:49:45,480 --> 00:49:47,400
the code does what it claims to do.

1461
00:49:47,400 --> 00:49:50,320
Makers can still iterate and test against realistic data,

1462
00:49:50,320 --> 00:49:52,480
but the pace slows down.

1463
00:49:52,480 --> 00:49:54,320
Before a skill moves to the next stage,

1464
00:49:54,320 --> 00:49:56,880
a lead must sign off to confirm they understand the code

1465
00:49:56,880 --> 00:49:58,360
and it is ready for validation.

1466
00:49:58,360 --> 00:50:00,240
Red zones are your production environments

1467
00:50:00,240 --> 00:50:02,760
where only certified skills are allowed to live.

1468
00:50:02,760 --> 00:50:06,000
Certification means the skill has passed security reviews,

1469
00:50:06,000 --> 00:50:08,720
performance tests and operational readiness checks.

1470
00:50:08,720 --> 00:50:10,160
It has been signed by an approver

1471
00:50:10,160 --> 00:50:12,920
and logged in a corporate catalog with clear ownership.

1472
00:50:12,920 --> 00:50:15,440
In these red zones, the mentality shifts entirely

1473
00:50:15,440 --> 00:50:17,160
toward minimizing surprises.

1474
00:50:17,160 --> 00:50:19,280
You block entire categories of connectors

1475
00:50:19,280 --> 00:50:23,240
and restrict network access to ensure nothing unexpected happens.

1476
00:50:23,240 --> 00:50:26,360
DLP is the enforcement layer that makes these zones real

1477
00:50:26,360 --> 00:50:27,920
rather than just a suggestion.

1478
00:50:27,920 --> 00:50:29,640
Data loss prevention policies govern

1479
00:50:29,640 --> 00:50:33,200
which connectors can talk to specific endpoints across your tenant.

1480
00:50:33,200 --> 00:50:35,760
You might block all unauthenticated HTTP connectors

1481
00:50:35,760 --> 00:50:38,120
at the top level, but allow them in green zones

1482
00:50:38,120 --> 00:50:39,480
so developers can learn.

1483
00:50:39,480 --> 00:50:41,240
As a skill moves toward production,

1484
00:50:41,240 --> 00:50:43,600
the policy becomes progressively more restrictive

1485
00:50:43,600 --> 00:50:46,120
until only allow listed endpoints are accessible.

1486
00:50:46,120 --> 00:50:48,680
Environment level isolation is the physical mechanism

1487
00:50:48,680 --> 00:50:50,720
for moving skills through these zones.

1488
00:50:50,720 --> 00:50:53,560
A skill does not just drift from green to yellow on its own.

1489
00:50:53,560 --> 00:50:56,320
A developer has to export it and request a promotion

1490
00:50:56,320 --> 00:50:58,120
which serves as your natural checkpoint

1491
00:50:58,120 --> 00:50:59,480
to ask if it was tested.

1492
00:50:59,480 --> 00:51:01,160
The skill travels through your environments

1493
00:51:01,160 --> 00:51:02,480
as a managed artifact,

1494
00:51:02,480 --> 00:51:04,200
ensuring that what ends up in production

1495
00:51:04,200 --> 00:51:05,560
is exactly what was approved.

1496
00:51:05,560 --> 00:51:07,440
The center of excellence plays a critical role here

1497
00:51:07,440 --> 00:51:10,000
by setting the baseline policies for the whole organization.

1498
00:51:10,000 --> 00:51:11,720
They decide which connectors are blocked

1499
00:51:11,720 --> 00:51:13,520
and which skills need formal approval

1500
00:51:13,520 --> 00:51:16,240
while domain-specific teams in HR or finance

1501
00:51:16,240 --> 00:51:18,800
maintain their own catalogs within those guardrails.

1502
00:51:18,800 --> 00:51:19,880
When you set it up this way,

1503
00:51:19,880 --> 00:51:21,760
governance is no longer just overhead,

1504
00:51:21,760 --> 00:51:24,120
but the actual operating system for your agents.

1505
00:51:24,120 --> 00:51:26,600
Trust and review patterns.

1506
00:51:26,600 --> 00:51:28,960
The Zoned governance framework gives you the structure

1507
00:51:28,960 --> 00:51:31,600
but structure without a process is just bureaucracy.

1508
00:51:31,600 --> 00:51:33,880
The actual work happens in how you review skills

1509
00:51:33,880 --> 00:51:35,680
before they enter yellow and red zones

1510
00:51:35,680 --> 00:51:37,240
and how you maintain that trust over time.

1511
00:51:37,240 --> 00:51:39,280
Start by treating external skills exactly

1512
00:51:39,280 --> 00:51:41,080
like you would treat third-party code

1513
00:51:41,080 --> 00:51:42,640
you are about to put into production.

1514
00:51:42,640 --> 00:51:45,160
You would not download a random library from GitHub

1515
00:51:45,160 --> 00:51:47,840
and drop it into your code base without looking at it first.

1516
00:51:47,840 --> 00:51:49,160
You would run security scans.

1517
00:51:49,160 --> 00:51:50,160
You would review the code,

1518
00:51:50,160 --> 00:51:52,280
you would test how it behaves in isolation

1519
00:51:52,280 --> 00:51:54,120
to understand what it accesses.

1520
00:51:54,120 --> 00:51:55,600
The same discipline applies here.

1521
00:51:55,600 --> 00:51:57,960
The review checklist is your operational standard.

1522
00:51:57,960 --> 00:52:00,240
Before a skill moves from green to yellow,

1523
00:52:00,240 --> 00:52:02,880
someone competent has to verify the details.

1524
00:52:02,880 --> 00:52:04,000
First is provenance.

1525
00:52:04,000 --> 00:52:05,440
Where did the skill come from?

1526
00:52:05,440 --> 00:52:06,680
You need to know who authored it

1527
00:52:06,680 --> 00:52:08,760
and if they have a track record you can trust.

1528
00:52:08,760 --> 00:52:11,040
A skill from an official Microsoft catalog

1529
00:52:11,040 --> 00:52:12,480
carries a different weight than one

1530
00:52:12,480 --> 00:52:14,120
from an anonymous GitHub user.

1531
00:52:14,120 --> 00:52:15,000
Next is intent.

1532
00:52:15,000 --> 00:52:16,440
What does the skill claim to do?

1533
00:52:16,440 --> 00:52:18,080
Read the metadata and the instructions

1534
00:52:18,080 --> 00:52:20,840
to see if the scope is narrow or dangerously broad.

1535
00:52:20,840 --> 00:52:22,040
Then look at dependencies.

1536
00:52:22,040 --> 00:52:23,880
What does the skill need to run?

1537
00:52:23,880 --> 00:52:25,560
You need to know which systems it calls

1538
00:52:25,560 --> 00:52:27,160
and what permissions it requires.

1539
00:52:27,160 --> 00:52:30,040
A skill that only needs read only access to customer data

1540
00:52:30,040 --> 00:52:32,320
is a much lower risk than one that needs right access

1541
00:52:32,320 --> 00:52:33,840
to your financial system.

1542
00:52:33,840 --> 00:52:35,640
Finally, check for injection risks.

1543
00:52:35,640 --> 00:52:37,720
Look for embedded instructions that could be overwritten

1544
00:52:37,720 --> 00:52:38,760
by a user.

1545
00:52:38,760 --> 00:52:41,440
If it passes untrusted input without validation,

1546
00:52:41,440 --> 00:52:43,720
an attacker could manipulate it into doing something

1547
00:52:43,720 --> 00:52:45,360
other than its stated purpose.

1548
00:52:45,360 --> 00:52:47,200
Static analysis is your first gate.

1549
00:52:47,200 --> 00:52:49,320
You run automated tools over the bundled scripts

1550
00:52:49,320 --> 00:52:53,080
to catch obvious red flags like hard-coded API keys or passwords.

1551
00:52:53,080 --> 00:52:55,600
These tools scan for suspicious file operations

1552
00:52:55,600 --> 00:52:57,960
and network calls to unexpected destinations.

1553
00:52:57,960 --> 00:52:59,920
It is fast and catches the easy mistakes,

1554
00:52:59,920 --> 00:53:01,440
but it is not enough on its own.

1555
00:53:01,440 --> 00:53:03,320
Static analysis finds known patterns,

1556
00:53:03,320 --> 00:53:05,120
but it misses new types of attacks.

1557
00:53:05,120 --> 00:53:08,160
Sandbox execution is where you see what the skill actually does.

1558
00:53:08,160 --> 00:53:10,840
You take the script and put it in an isolated environment

1559
00:53:10,840 --> 00:53:12,800
like a container or a virtual machine.

1560
00:53:12,800 --> 00:53:14,960
You give it minimal permissions and then you watch.

1561
00:53:14,960 --> 00:53:15,880
Does it read files?

1562
00:53:15,880 --> 00:53:16,640
Which ones?

1563
00:53:16,640 --> 00:53:19,760
Does it try to start new processes or change the system state?

1564
00:53:19,760 --> 00:53:20,960
You capture everything it does

1565
00:53:20,960 --> 00:53:23,600
and compare those observations to the original intent.

1566
00:53:23,600 --> 00:53:25,440
If a skill says it validates a text ID,

1567
00:53:25,440 --> 00:53:28,720
but tries to make external HTTP calls, you have a mismatch.

1568
00:53:28,720 --> 00:53:30,840
You investigate that before you ever deploy it.

1569
00:53:30,840 --> 00:53:32,680
Permission scoping is how you limit the damage

1570
00:53:32,680 --> 00:53:33,800
when something goes wrong.

1571
00:53:33,800 --> 00:53:36,280
If a skill needs to call an approval API,

1572
00:53:36,280 --> 00:53:39,040
it does not need access to your entire database.

1573
00:53:39,040 --> 00:53:40,560
You create a service account with permissions

1574
00:53:40,560 --> 00:53:42,760
limited to that one specific operation.

1575
00:53:42,760 --> 00:53:45,000
By giving the skill those exact credentials,

1576
00:53:45,000 --> 00:53:47,200
you are following the rule of least privilege.

1577
00:53:47,200 --> 00:53:48,720
If the skill is compromised,

1578
00:53:48,720 --> 00:53:51,520
the attacker can only do what that one credential allows.

1579
00:53:51,520 --> 00:53:54,200
The recertification requirement reflects a hard truth.

1580
00:53:54,200 --> 00:53:56,680
Skills drift, an approved skill is frozen in time

1581
00:53:56,680 --> 00:53:58,160
the moment you say yes to it,

1582
00:53:58,160 --> 00:54:00,200
but systems change in APIs evolve.

1583
00:54:00,200 --> 00:54:02,320
A skill written for an older version of a database

1584
00:54:02,320 --> 00:54:04,280
might not work correctly on the new one.

1585
00:54:04,280 --> 00:54:06,560
Even worse, a new vulnerability in a library

1586
00:54:06,560 --> 00:54:08,360
can make a safe skill risky overnight.

1587
00:54:08,360 --> 00:54:10,880
Reviews are not one-time events, they are checkpoints.

1588
00:54:10,880 --> 00:54:13,880
Every quarter, you should reexamine your high-risk skills

1589
00:54:13,880 --> 00:54:15,920
to see if the threat landscape has changed.

1590
00:54:15,920 --> 00:54:17,840
This discipline costs resources.

1591
00:54:17,840 --> 00:54:19,680
Reviewing takes time and sandbox testing

1592
00:54:19,680 --> 00:54:21,200
requires infrastructure.

1593
00:54:21,200 --> 00:54:24,000
Recertification is an ongoing cost you have to pay.

1594
00:54:24,000 --> 00:54:26,600
But this is the price of running untrusted codes safely.

1595
00:54:26,600 --> 00:54:29,520
Organizations that skip these steps get a little speed now.

1596
00:54:29,520 --> 00:54:31,360
But they get a lot of vulnerability later.

1597
00:54:31,360 --> 00:54:33,800
The teams that build these practices into their daily flow

1598
00:54:33,800 --> 00:54:36,720
are the ones who make skills safe to use at scale.

1599
00:54:36,720 --> 00:54:38,520
The skill catalog future.

1600
00:54:38,520 --> 00:54:40,720
Today, the way skills work in co-pilot studio

1601
00:54:40,720 --> 00:54:42,000
has a big limitation.

1602
00:54:42,000 --> 00:54:44,480
Skills are stuck inside the agent that contains them.

1603
00:54:44,480 --> 00:54:47,360
If you build a skill for agent A and want it in agent B,

1604
00:54:47,360 --> 00:54:49,560
you have to duplicate it or refactor everything

1605
00:54:49,560 --> 00:54:51,560
to use a shared dataverse skill.

1606
00:54:51,560 --> 00:54:53,320
Duplication is a mess to maintain.

1607
00:54:53,320 --> 00:54:55,680
Refactoring works, but it takes a lot of extra effort.

1608
00:54:55,680 --> 00:54:56,960
Neither way is easy.

1609
00:54:56,960 --> 00:54:58,480
But this constraint is temporary.

1610
00:54:58,480 --> 00:55:01,200
It shows where the market is right now in mid-2026,

1611
00:55:01,200 --> 00:55:02,400
not where it is going.

1612
00:55:02,400 --> 00:55:03,680
The path forward is clear.

1613
00:55:03,680 --> 00:55:06,360
Microsoft is moving toward tenant-wide skill catalogs.

1614
00:55:06,360 --> 00:55:08,080
This means you author a skill once,

1615
00:55:08,080 --> 00:55:11,240
govern it once, and use it across every agent in your company.

1616
00:55:11,240 --> 00:55:12,800
GitHub co-pilot is doing the same thing

1617
00:55:12,800 --> 00:55:14,160
with portable packages.

1618
00:55:14,160 --> 00:55:15,960
Claude code is following along, too.

1619
00:55:15,960 --> 00:55:19,480
Everything is moving toward the openskill.md standard

1620
00:55:19,480 --> 00:55:22,320
as the common language for what an agent can do.

1621
00:55:22,320 --> 00:55:24,600
The goal is to write once and run everywhere.

1622
00:55:24,600 --> 00:55:27,040
You might author a skill for onboarding a new vendor.

1623
00:55:27,040 --> 00:55:29,520
That skill is just instructions and metadata in markdown,

1624
00:55:29,520 --> 00:55:31,440
maybe bundled with a few scripts.

1625
00:55:31,440 --> 00:55:33,680
That same skill can run in co-pilot studio,

1626
00:55:33,680 --> 00:55:35,400
in GitHub co-pilot for automation,

1627
00:55:35,400 --> 00:55:36,880
or in Claude code for engineering.

1628
00:55:36,880 --> 00:55:39,320
It becomes one source of truth with many different ways

1629
00:55:39,320 --> 00:55:40,040
to use it.

1630
00:55:40,040 --> 00:55:43,240
You do not have to rewrite it or adapt it for every new tool.

1631
00:55:43,240 --> 00:55:44,920
The skill carries its intent with it.

1632
00:55:44,920 --> 00:55:46,160
This is not just a theory.

1633
00:55:46,160 --> 00:55:47,800
Organizations are already building this way

1634
00:55:47,800 --> 00:55:49,960
by keeping their skills and git repositories.

1635
00:55:49,960 --> 00:55:51,360
They version them just like code

1636
00:55:51,360 --> 00:55:53,320
and move them across different platforms.

1637
00:55:53,320 --> 00:55:54,680
The tools are still a bit clunky

1638
00:55:54,680 --> 00:55:56,640
and you might have to tweak a format here and there,

1639
00:55:56,640 --> 00:55:57,560
but the vision is set.

1640
00:55:57,560 --> 00:56:00,800
The skill is becoming the portable unit of what an agent can do.

1641
00:56:00,800 --> 00:56:03,600
Dataverse intelligence is the layer that makes this vision work.

1642
00:56:03,600 --> 00:56:05,880
Business skills stored in Dataverse are company assets

1643
00:56:05,880 --> 00:56:08,880
that you own, approve, and eventually retire.

1644
00:56:08,880 --> 00:56:10,280
They are not locked into one agent.

1645
00:56:10,280 --> 00:56:12,080
You publish them to an enterprise library

1646
00:56:12,080 --> 00:56:14,120
where any authorized agent can find them.

1647
00:56:14,120 --> 00:56:16,800
In this model, the Dataverse skill is the system of record.

1648
00:56:16,800 --> 00:56:18,240
And the agents are just clients.

1649
00:56:18,240 --> 00:56:19,800
This flips the old architecture

1650
00:56:19,800 --> 00:56:22,400
where every agent had to own its own knowledge.

1651
00:56:22,400 --> 00:56:24,240
The marketplace for this is still forming,

1652
00:56:24,240 --> 00:56:25,880
but the potential is huge.

1653
00:56:25,880 --> 00:56:28,320
Eventually vendors will publish pre-built skill libraries

1654
00:56:28,320 --> 00:56:29,600
for specific jobs.

1655
00:56:29,600 --> 00:56:32,160
Sales force could offer a library for CRM tasks

1656
00:56:32,160 --> 00:56:34,880
and SAP could do the same for financial processes.

1657
00:56:34,880 --> 00:56:36,680
Your company would subscribe to these libraries

1658
00:56:36,680 --> 00:56:37,920
and immediately give your agents

1659
00:56:37,920 --> 00:56:39,720
the best practices for those fields.

1660
00:56:39,720 --> 00:56:41,440
Right now that is still a bit speculative,

1661
00:56:41,440 --> 00:56:44,280
the ecosystem is young and the quality of skills varies a lot.

1662
00:56:44,280 --> 00:56:45,320
We do not have solid models

1663
00:56:45,320 --> 00:56:47,520
for how companies will sell or distribute these yet,

1664
00:56:47,520 --> 00:56:49,040
but the architecture is ready for it

1665
00:56:49,040 --> 00:56:50,800
and the pressure to build it is real.

1666
00:56:50,800 --> 00:56:52,880
Every big company asking how to scale their agents

1667
00:56:52,880 --> 00:56:53,760
hits the same wall.

1668
00:56:53,760 --> 00:56:55,880
They need to know how to share expertise efficiently.

1669
00:56:55,880 --> 00:56:56,920
Skills are the answer

1670
00:56:56,920 --> 00:57:00,160
and the infrastructure to share them is finally showing up.

1671
00:57:00,160 --> 00:57:02,400
In 2026, an enterprise skill library

1672
00:57:02,400 --> 00:57:05,560
is a curated collection of procedures kept in a central system.

1673
00:57:05,560 --> 00:57:07,080
It is the company's version of a manual

1674
00:57:07,080 --> 00:57:08,800
but the manual is machine readable

1675
00:57:08,800 --> 00:57:10,440
so agents can actually do the work.

1676
00:57:10,440 --> 00:57:12,000
Domain teams keep it updated.

1677
00:57:12,000 --> 00:57:13,360
It is auditable and compliant.

1678
00:57:13,360 --> 00:57:14,680
It is the foundation for how

1679
00:57:14,680 --> 00:57:17,000
an AI native company actually functions.

1680
00:57:17,000 --> 00:57:18,800
Most organizations are not there yet.

1681
00:57:18,800 --> 00:57:20,160
They are still building single agents

1682
00:57:20,160 --> 00:57:22,400
or just learning how skills work in one place,

1683
00:57:22,400 --> 00:57:23,800
but you can see where this is going.

1684
00:57:23,800 --> 00:57:25,360
First comes skill sharing,

1685
00:57:25,360 --> 00:57:27,240
then using them across different platforms

1686
00:57:27,240 --> 00:57:29,160
and finally a full marketplace.

1687
00:57:29,160 --> 00:57:30,960
Each step builds on the one before it.

1688
00:57:30,960 --> 00:57:32,560
The zoned governance framework

1689
00:57:32,560 --> 00:57:35,080
actually becomes more important as you share more.

1690
00:57:35,080 --> 00:57:38,040
You need stronger controls when one skill serves 50 agents

1691
00:57:38,040 --> 00:57:39,560
instead of just one.

1692
00:57:39,560 --> 00:57:41,880
The dataverse endpoint becomes your central hub

1693
00:57:41,880 --> 00:57:44,120
for everything your company knows how to do.

1694
00:57:44,120 --> 00:57:45,880
This is where the architecture gets deep.

1695
00:57:45,880 --> 00:57:47,680
Skills are how you encode intelligence.

1696
00:57:47,680 --> 00:57:49,240
You do not put it in a separate document

1697
00:57:49,240 --> 00:57:50,440
or a static system.

1698
00:57:50,440 --> 00:57:52,160
You turn it into an executable procedure

1699
00:57:52,160 --> 00:57:53,480
that an agent can use.

1700
00:57:53,480 --> 00:57:56,320
The skill catalog becomes the memory of your organization.

1701
00:57:56,320 --> 00:57:58,240
It is how you make sure the right way to do things

1702
00:57:58,240 --> 00:58:00,080
is shared, improved and enforced

1703
00:58:00,080 --> 00:58:01,760
across every agent you run.

1704
00:58:01,760 --> 00:58:04,560
The orchestration shift, your old chatbot was built on a map.

1705
00:58:04,560 --> 00:58:06,360
It used deterministic dialog trees.

1706
00:58:06,360 --> 00:58:08,760
If the user clicks A, the bot goes to B.

1707
00:58:08,760 --> 00:58:10,920
If the user says X, the bot follows path Y,

1708
00:58:10,920 --> 00:58:12,560
it was rigid, it was predictable

1709
00:58:12,560 --> 00:58:14,680
and it was completely incapable of judgment.

1710
00:58:14,680 --> 00:58:17,440
The system didn't think it just navigated a pre-drawn

1711
00:58:17,440 --> 00:58:19,320
maze, modern agents have moved away from that.

1712
00:58:19,320 --> 00:58:20,680
They don't follow a map anymore.

1713
00:58:20,680 --> 00:58:21,920
They use semantic routing.

1714
00:58:21,920 --> 00:58:23,800
When a user asks a question, the orchestrator

1715
00:58:23,800 --> 00:58:25,520
doesn't look for a hard-coded branch.

1716
00:58:25,520 --> 00:58:27,280
It looks for meaning, it asks,

1717
00:58:27,280 --> 00:58:29,520
what is this person actually trying to do?

1718
00:58:29,520 --> 00:58:32,120
It asks, which skill in my library is the best match

1719
00:58:32,120 --> 00:58:33,480
for this specific moment?

1720
00:58:33,480 --> 00:58:35,720
It asks, do I even need a skill?

1721
00:58:35,720 --> 00:58:37,880
Or can I handle this with my own reasoning?

1722
00:58:37,880 --> 00:58:39,240
This isn't treat reversal.

1723
00:58:39,240 --> 00:58:41,760
It's dynamic matching between a user's intent

1724
00:58:41,760 --> 00:58:43,480
and the agent's capabilities.

1725
00:58:43,480 --> 00:58:45,440
Here is how that decision actually happens.

1726
00:58:45,440 --> 00:58:46,960
A message comes in, the orchestrator

1727
00:58:46,960 --> 00:58:49,520
looks at the metadata of every available skill.

1728
00:58:49,520 --> 00:58:52,360
The names, the descriptions, the purpose, then it scores them.

1729
00:58:52,360 --> 00:58:54,440
It might decide there is an 85% match

1730
00:58:54,440 --> 00:58:56,080
for the leave eligibility skill,

1731
00:58:56,080 --> 00:58:59,080
but only a 5% match for benefits, the high score wins,

1732
00:58:59,080 --> 00:59:00,160
the skill loads.

1733
00:59:00,160 --> 00:59:01,280
But here is the shift.

1734
00:59:01,280 --> 00:59:03,000
The skill doesn't lock the agent into a cage.

1735
00:59:03,000 --> 00:59:05,080
It provides expertise, not a script.

1736
00:59:05,080 --> 00:59:07,640
It says, if you're talking about leave,

1737
00:59:07,640 --> 00:59:10,120
here are the rules and the steps you need to follow.

1738
00:59:10,120 --> 00:59:11,840
But the model still has its own brain.

1739
00:59:11,840 --> 00:59:13,960
If the user has a weird, specific nuance

1740
00:59:13,960 --> 00:59:16,360
that the skill didn't plan for, the model adapts

1741
00:59:16,360 --> 00:59:18,400
if the request is slightly outside the box,

1742
00:59:18,400 --> 00:59:21,120
the model extrapolates, the skill orients the reasoning.

1743
00:59:21,120 --> 00:59:22,680
But it doesn't constrain the output.

1744
00:59:22,680 --> 00:59:24,760
You see this most clearly in multi-turn conversations.

1745
00:59:24,760 --> 00:59:27,280
A user asks about leave, the leave skill loads,

1746
00:59:27,280 --> 00:59:29,640
the agent answers, then the user follows up.

1747
00:59:29,640 --> 00:59:31,080
What if I want to extend it?

1748
00:59:31,080 --> 00:59:33,840
The orchestrator realizes we are still talking about leave.

1749
00:59:33,840 --> 00:59:36,120
The skill stays loaded, the context remains active,

1750
00:59:36,120 --> 00:59:37,760
because the skill is already there,

1751
00:59:37,760 --> 00:59:39,600
the cost of thinking stays low.

1752
00:59:39,600 --> 00:59:41,160
The agent doesn't have to relearn the rules

1753
00:59:41,160 --> 00:59:43,840
for every new sentence, but then the user pivots.

1754
00:59:43,840 --> 00:59:45,440
How do I update my direct deposit?

1755
00:59:45,440 --> 00:59:47,800
That isn't leave, that's payroll.

1756
00:59:47,800 --> 00:59:49,280
The orchestrator sees the mismatch

1757
00:59:49,280 --> 00:59:51,360
and unloads the leave skill immediately.

1758
00:59:51,360 --> 00:59:54,360
It shifts its cognitive load to a new set of expertise.

1759
00:59:54,360 --> 00:59:56,160
The path changes because the intent changed,

1760
00:59:56,160 --> 00:59:58,000
sometimes these skills chain together.

1761
00:59:58,000 --> 00:59:59,600
One skill validates a vendor,

1762
00:59:59,600 --> 01:00:01,400
and that success triggers a second skill

1763
01:00:01,400 --> 01:00:02,840
to set up payment terms.

1764
01:00:02,840 --> 01:00:04,280
The completion of that might trigger

1765
01:00:04,280 --> 01:00:06,400
a third skill to notify procurement.

1766
01:00:06,400 --> 01:00:08,200
Work flows flow, but the orchestrator

1767
01:00:08,200 --> 01:00:09,680
is the one managing the timing.

1768
01:00:09,680 --> 01:00:11,520
This is why modern agents feel different.

1769
01:00:11,520 --> 01:00:13,040
The skill provides the expertise,

1770
01:00:13,040 --> 01:00:14,560
the model provides the judgment.

1771
01:00:14,560 --> 01:00:17,240
The model without a skill is slow and makes mistakes.

1772
01:00:17,240 --> 01:00:19,920
A skill without a model is just a rigid brittle procedure.

1773
01:00:19,920 --> 01:00:20,960
But when you combine them,

1774
01:00:20,960 --> 01:00:23,360
you get a system that knows exactly when to follow the rules

1775
01:00:23,360 --> 01:00:24,840
and when to use its head.

1776
01:00:24,840 --> 01:00:26,400
Measuring skill effectiveness,

1777
01:00:26,400 --> 01:00:28,680
there is a massive gap between a skill that should work,

1778
01:00:28,680 --> 01:00:30,640
and a skill that actually works in production,

1779
01:00:30,640 --> 01:00:32,280
most organizations stumble here.

1780
01:00:32,280 --> 01:00:34,320
They build, they deploy, and then they assume

1781
01:00:34,320 --> 01:00:35,400
everything is fine.

1782
01:00:35,400 --> 01:00:37,120
Months go by, the agent feels slow,

1783
01:00:37,120 --> 01:00:38,200
the savings aren't there,

1784
01:00:38,200 --> 01:00:39,640
the team stops maintaining the skills

1785
01:00:39,640 --> 01:00:41,800
because nobody knows what they're actually doing.

1786
01:00:41,800 --> 01:00:43,960
The reason is simple, nobody measured.

1787
01:00:43,960 --> 01:00:46,000
If you want this to work, you have to be systematic.

1788
01:00:46,000 --> 01:00:48,240
There are five dimensions that actually matter.

1789
01:00:48,240 --> 01:00:49,960
First is triggering accuracy.

1790
01:00:49,960 --> 01:00:52,720
Does the orchestrator call the skill when it's supposed to?

1791
01:00:52,720 --> 01:00:53,960
If your leave skill is firing

1792
01:00:53,960 --> 01:00:55,560
when people ask about their paycheck,

1793
01:00:55,560 --> 01:00:56,600
your routing is broken.

1794
01:00:56,600 --> 01:00:57,680
You need to watch the rates.

1795
01:00:57,680 --> 01:00:59,560
If a skill usually fires 10 times a day

1796
01:00:59,560 --> 01:01:01,160
and suddenly drops to two.

1797
01:01:01,160 --> 01:01:02,280
Something changed.

1798
01:01:02,280 --> 01:01:04,120
Either your users stopped caring

1799
01:01:04,120 --> 01:01:06,040
or your routing quality degraded.

1800
01:01:06,040 --> 01:01:07,280
Second is isolation.

1801
01:01:07,280 --> 01:01:09,440
Does the skill work when it's the only thing running?

1802
01:01:09,440 --> 01:01:11,520
Don't worry about complex interactions yet.

1803
01:01:11,520 --> 01:01:14,360
Just run core scenarios against the skill by itself

1804
01:01:14,360 --> 01:01:16,000
to make sure the logic holds up.

1805
01:01:16,000 --> 01:01:17,160
Third is coexistence.

1806
01:01:17,160 --> 01:01:18,680
This is where things get messy.

1807
01:01:18,680 --> 01:01:20,800
What happens when you load five skills at once?

1808
01:01:20,800 --> 01:01:22,360
Do the descriptions overlap so much

1809
01:01:22,360 --> 01:01:24,200
that the orchestrator gets confused?

1810
01:01:24,200 --> 01:01:26,840
Do the instructions in skill A, contradict skill B?

1811
01:01:26,840 --> 01:01:28,840
Testing coexistence ensures your skills

1812
01:01:28,840 --> 01:01:30,960
aren't fighting each other for control.

1813
01:01:30,960 --> 01:01:32,680
Fourth is instruction following.

1814
01:01:32,680 --> 01:01:33,880
When the skill is active,

1815
01:01:33,880 --> 01:01:35,560
does the model actually listen to it?

1816
01:01:35,560 --> 01:01:37,040
You have to watch the reasoning path.

1817
01:01:37,040 --> 01:01:38,560
If the skill says a step is mandatory

1818
01:01:38,560 --> 01:01:41,040
and the model skips it, you have a design flaw.

1819
01:01:41,040 --> 01:01:42,360
Usually this means your instructions

1820
01:01:42,360 --> 01:01:45,000
are too ambiguous for the model to follow.

1821
01:01:45,000 --> 01:01:46,520
Fifth is output quality.

1822
01:01:46,520 --> 01:01:47,680
This is the bottom line.

1823
01:01:47,680 --> 01:01:48,720
The routing worked.

1824
01:01:48,720 --> 01:01:50,240
The instructions were followed.

1825
01:01:50,240 --> 01:01:51,560
But was the answer right?

1826
01:01:51,560 --> 01:01:53,760
If a human expert wouldn't sign off on the response,

1827
01:01:53,760 --> 01:01:54,760
the skill failed.

1828
01:01:54,760 --> 01:01:56,600
To make this real, you have to benchmark.

1829
01:01:56,600 --> 01:01:58,440
Before you turn a workflow into a skill,

1830
01:01:58,440 --> 01:01:59,480
you need a baseline.

1831
01:01:59,480 --> 01:02:00,760
Take 10 common requests.

1832
01:02:00,760 --> 01:02:02,360
Run them through your old system.

1833
01:02:02,360 --> 01:02:05,280
Record the success rate, the token count, the latency,

1834
01:02:05,280 --> 01:02:06,920
and how many turns it took to finish.

1835
01:02:06,920 --> 01:02:10,400
Then, deploy the skill and run those same 10 requests again.

1836
01:02:10,400 --> 01:02:11,440
Compare the numbers.

1837
01:02:11,440 --> 01:02:12,960
If the success rate didn't go up

1838
01:02:12,960 --> 01:02:14,720
or the token count didn't go down,

1839
01:02:14,720 --> 01:02:15,640
the skill didn't help.

1840
01:02:15,640 --> 01:02:16,400
Don't guess.

1841
01:02:16,400 --> 01:02:17,440
Measure.

1842
01:02:17,440 --> 01:02:19,480
Real world testing is the final hurdle.

1843
01:02:19,480 --> 01:02:21,720
Lab tests are clean, but users are messy.

1844
01:02:21,720 --> 01:02:23,240
They ask things you didn't expect.

1845
01:02:23,240 --> 01:02:25,000
They use language you didn't plan for.

1846
01:02:25,000 --> 01:02:25,920
Collect the logs.

1847
01:02:25,920 --> 01:02:27,120
Look for the failure modes.

1848
01:02:27,120 --> 01:02:28,600
Maybe an overbroad description

1849
01:02:28,600 --> 01:02:30,560
is making the skill fire too often.

1850
01:02:30,560 --> 01:02:32,520
Maybe an API changed and your bundled script

1851
01:02:32,520 --> 01:02:34,040
is now spitting out garbage.

1852
01:02:34,040 --> 01:02:35,880
Maybe two skills are now locked in a conflict

1853
01:02:35,880 --> 01:02:37,080
you didn't see in the lab.

1854
01:02:37,080 --> 01:02:39,360
When the data shows a skill is missing its targets,

1855
01:02:39,360 --> 01:02:40,200
you refactor it.

1856
01:02:40,200 --> 01:02:41,920
If it's trying to do too much, you split it.

1857
01:02:41,920 --> 01:02:44,200
If it's not being used, you retire it.

1858
01:02:44,200 --> 01:02:45,800
Let the data drive the architecture.

1859
01:02:45,800 --> 01:02:46,960
Not your assumptions.

1860
01:02:46,960 --> 01:02:48,360
Implementation patterns.

1861
01:02:48,360 --> 01:02:49,960
Where do you actually start when you're ready

1862
01:02:49,960 --> 01:02:51,320
to build your first skill?

1863
01:02:51,320 --> 01:02:53,640
Most teams make this harder than it needs to be.

1864
01:02:53,640 --> 01:02:56,040
They spend weeks designing a perfect modular architecture

1865
01:02:56,040 --> 01:02:57,720
or planning 30 skills in advance.

1866
01:02:57,720 --> 01:02:59,880
They create taxonomies and governance frameworks

1867
01:02:59,880 --> 01:03:01,760
before they've even written a single line of code.

1868
01:03:01,760 --> 01:03:03,280
That's organizational theater.

1869
01:03:03,280 --> 01:03:04,120
And it doesn't help.

1870
01:03:04,120 --> 01:03:05,160
You need to start small.

1871
01:03:05,160 --> 01:03:06,360
Pick one workflow.

1872
01:03:06,360 --> 01:03:07,600
Look for something high frequency

1873
01:03:07,600 --> 01:03:09,520
that your agents handle repeatedly.

1874
01:03:09,520 --> 01:03:12,480
It should be a task where you know every single step by heart.

1875
01:03:12,480 --> 01:03:14,600
Maybe it's a password reset for your IT agent

1876
01:03:14,600 --> 01:03:16,600
or an onboarding checklist for HR.

1877
01:03:16,600 --> 01:03:18,920
It could even be a purchase or a validation

1878
01:03:18,920 --> 01:03:20,200
for a finance copilot.

1879
01:03:20,200 --> 01:03:23,280
Just find one task that users ask about 10 or 20 times a week.

1880
01:03:23,280 --> 01:03:24,280
That is your candidate.

1881
01:03:24,280 --> 01:03:25,320
Extract the logic.

1882
01:03:25,320 --> 01:03:27,480
But don't try to redesign or optimize it yet.

1883
01:03:27,480 --> 01:03:29,680
Just move what you already have into a skill.

1884
01:03:29,680 --> 01:03:32,080
You likely have instructions sitting in a document,

1885
01:03:32,080 --> 01:03:34,440
a wiki or an old email that nobody reads.

1886
01:03:34,440 --> 01:03:35,680
That is your starting material.

1887
01:03:35,680 --> 01:03:37,280
Write a skilled MD file

1888
01:03:37,280 --> 01:03:38,440
and give it a clear name.

1889
01:03:38,440 --> 01:03:39,960
Write the instructions in markdown

1890
01:03:39,960 --> 01:03:42,160
and bundle any templates or reference data.

1891
01:03:42,160 --> 01:03:43,200
It won't be perfect.

1892
01:03:43,200 --> 01:03:44,400
And that's fine.

1893
01:03:44,400 --> 01:03:45,720
Perfection comes later.

1894
01:03:45,720 --> 01:03:49,160
The naming convention matters more than you think.

1895
01:03:49,160 --> 01:03:50,720
Vague names cause rooting failures

1896
01:03:50,720 --> 01:03:52,840
because they confuse the orchestrator.

1897
01:03:52,840 --> 01:03:55,000
A name like HR help is too broad.

1898
01:03:55,000 --> 01:03:57,360
So it triggers on everything HR related

1899
01:03:57,360 --> 01:03:59,280
even when you need something specific.

1900
01:03:59,280 --> 01:04:01,440
HR leave eligibility triage is much better

1901
01:04:01,440 --> 01:04:02,680
because it's precise.

1902
01:04:02,680 --> 01:04:04,600
The pattern that actually works is domain

1903
01:04:04,600 --> 01:04:06,400
then specific task then constrained.

1904
01:04:06,400 --> 01:04:07,920
The domain narrows the field,

1905
01:04:07,920 --> 01:04:09,680
the task clarifies the action

1906
01:04:09,680 --> 01:04:11,640
and the constraint marks the boundary.

1907
01:04:11,640 --> 01:04:13,720
If you name a skill, finance, invoice approval

1908
01:04:13,720 --> 01:04:17,040
under $10,000, the orchestrator knows exactly when to use it.

1909
01:04:17,040 --> 01:04:18,560
Writing instructions for an orchestrator

1910
01:04:18,560 --> 01:04:20,000
is a specific discipline.

1911
01:04:20,000 --> 01:04:21,720
You aren't documenting for humans here.

1912
01:04:21,720 --> 01:04:23,560
You're explaining a process to a reasoning system

1913
01:04:23,560 --> 01:04:25,640
that needs to know exactly which decisions to make.

1914
01:04:25,640 --> 01:04:28,080
Use numbered steps and make every condition explicit.

1915
01:04:28,080 --> 01:04:30,680
Instead of saying high amounts need more review,

1916
01:04:30,680 --> 01:04:31,760
give it a clear rule.

1917
01:04:31,760 --> 01:04:34,520
Tell the system that if an amount is under 5,000,

1918
01:04:34,520 --> 01:04:35,720
it goes to level one.

1919
01:04:35,720 --> 01:04:38,280
If it's between 5,000 and 50,000, it goes to level two.

1920
01:04:38,280 --> 01:04:41,320
If it exceeds 50,000, it requires executive sign-off.

1921
01:04:41,320 --> 01:04:43,720
You need to be concrete, specific, and conditional.

1922
01:04:43,720 --> 01:04:46,280
The soft-pointing pattern is how you reference tools

1923
01:04:46,280 --> 01:04:49,080
without locking the skill to one specific piece of software.

1924
01:04:49,080 --> 01:04:51,080
Don't say call the approval API.

1925
01:04:51,080 --> 01:04:53,800
Instead, tell the system to root to the approval workflow

1926
01:04:53,800 --> 01:04:54,760
for this level.

1927
01:04:54,760 --> 01:04:57,240
You shouldn't embed specific API endpoints.

1928
01:04:57,240 --> 01:04:59,480
Reference logical capabilities instead.

1929
01:04:59,480 --> 01:05:01,080
Tell the skill to validate the vendor

1930
01:05:01,080 --> 01:05:04,160
against the tax ID service without specifying which service

1931
01:05:04,160 --> 01:05:05,240
or how to call it.

1932
01:05:05,240 --> 01:05:06,760
This allows the same skill to work

1933
01:05:06,760 --> 01:05:08,880
across different co-pilot studio setups.

1934
01:05:08,880 --> 01:05:10,400
One agent might have a built-in workflow

1935
01:05:10,400 --> 01:05:12,800
while another delegates to an external system.

1936
01:05:12,800 --> 01:05:14,640
The instructions work both ways because the reference

1937
01:05:14,640 --> 01:05:16,160
is logical, not technical.

1938
01:05:16,160 --> 01:05:17,920
Versioning is how you stop skills from turning

1939
01:05:17,920 --> 01:05:19,080
into technical debt.

1940
01:05:19,080 --> 01:05:21,680
You have to treat them like code, give them version numbers,

1941
01:05:21,680 --> 01:05:23,560
and document every change you make.

1942
01:05:23,560 --> 01:05:26,160
When you update a skill, increment that version number,

1943
01:05:26,160 --> 01:05:28,320
so you know exactly what's running in production.

1944
01:05:28,320 --> 01:05:30,000
If something breaks after an update,

1945
01:05:30,000 --> 01:05:32,360
you can just roll back to the previous version.

1946
01:05:32,360 --> 01:05:34,440
If you need to know why the behavior changed,

1947
01:05:34,440 --> 01:05:35,600
your change log will tell you,

1948
01:05:35,600 --> 01:05:37,560
this turns your skills into reliable assets

1949
01:05:37,560 --> 01:05:39,640
rather than mysterious black boxes.

1950
01:05:39,640 --> 01:05:41,280
The A-LIM integration is where the skill

1951
01:05:41,280 --> 01:05:43,120
becomes a real enterprise asset.

1952
01:05:43,120 --> 01:05:45,480
You aren't just saving a file on a hard drive.

1953
01:05:45,480 --> 01:05:48,000
You're adding it to a solution in the power platform

1954
01:05:48,000 --> 01:05:51,600
that moves through your dev, test, and production environments.

1955
01:05:51,600 --> 01:05:53,200
When you update a skill, that update

1956
01:05:53,200 --> 01:05:54,920
moves through your existing pipeline.

1957
01:05:54,920 --> 01:05:56,120
When you retire a skill,

1958
01:05:56,120 --> 01:05:58,320
it's removed from every environment at once.

1959
01:05:58,320 --> 01:05:59,680
This is version control and governance

1960
01:05:59,680 --> 01:06:01,160
built directly into the system,

1961
01:06:01,160 --> 01:06:02,640
not bolted on as an afterthought.

1962
01:06:02,640 --> 01:06:04,800
Implementation patterns are where theory finally

1963
01:06:04,800 --> 01:06:06,120
becomes operational.

1964
01:06:06,120 --> 01:06:07,720
You aren't just scaling one skill.

1965
01:06:07,720 --> 01:06:09,480
You're establishing a repeatable practice.

1966
01:06:09,480 --> 01:06:12,760
Start small, name things precisely, and write clearly.

1967
01:06:12,760 --> 01:06:14,560
If you do this with your first skill,

1968
01:06:14,560 --> 01:06:17,320
scaling to 1,000 others becomes a simple repetition

1969
01:06:17,320 --> 01:06:19,000
of the same process.

1970
01:06:19,000 --> 01:06:21,600
The strategic shift, the transformation happening right now

1971
01:06:21,600 --> 01:06:23,000
isn't just a small change.

1972
01:06:23,000 --> 01:06:24,000
It's architectural.

1973
01:06:24,000 --> 01:06:25,800
We are moving away from the old model

1974
01:06:25,800 --> 01:06:27,480
where you build one omniscient agent

1975
01:06:27,480 --> 01:06:29,920
that tries to know everything about your entire organization.

1976
01:06:29,920 --> 01:06:31,800
The new model is an agent that knows exactly

1977
01:06:31,800 --> 01:06:34,600
what it needs to know, only when it needs to know it.

1978
01:06:34,600 --> 01:06:36,200
That sounds like a small distinction,

1979
01:06:36,200 --> 01:06:37,600
but it changes everything.

1980
01:06:37,600 --> 01:06:39,480
The old model told us to build a chatbot

1981
01:06:39,480 --> 01:06:41,240
and give it every piece of knowledge possible.

1982
01:06:41,240 --> 01:06:42,920
We thought comprehensive men smart.

1983
01:06:42,920 --> 01:06:44,160
We thought it meant useful.

1984
01:06:44,160 --> 01:06:45,960
The agent would load everything up front

1985
01:06:45,960 --> 01:06:48,320
from HR policies and IT procedures

1986
01:06:48,320 --> 01:06:50,560
to finance workflows and compliance rules.

1987
01:06:50,560 --> 01:06:51,440
It was a polymath.

1988
01:06:51,440 --> 01:06:54,280
We assumed that if the agent read every manual cover to cover,

1989
01:06:54,280 --> 01:06:55,280
it could answer anything.

1990
01:06:55,280 --> 01:06:56,080
But here's the problem.

1991
01:06:56,080 --> 01:06:58,800
Omniscience is expensive, it's slow, it's prone to errors,

1992
01:06:58,800 --> 01:07:00,240
and it simply doesn't scale.

1993
01:07:00,240 --> 01:07:01,800
The new model is fundamentally different.

1994
01:07:01,800 --> 01:07:03,640
You build a focused agent that has the ability

1995
01:07:03,640 --> 01:07:05,240
to load expertise on demand.

1996
01:07:05,240 --> 01:07:07,200
It becomes selective about what it looks at.

1997
01:07:07,200 --> 01:07:09,960
An agent in this model doesn't pretend to know everything.

1998
01:07:09,960 --> 01:07:11,720
It knows a little bit about the whole company.

1999
01:07:11,720 --> 01:07:13,840
And a lot about what is relevant right now.

2000
01:07:13,840 --> 01:07:16,040
When a user asks about leave eligibility,

2001
01:07:16,040 --> 01:07:18,760
the agent doesn't waste time rummaging through payment systems

2002
01:07:18,760 --> 01:07:20,000
or compliance training.

2003
01:07:20,000 --> 01:07:22,320
It loads the leave expertise, period.

2004
01:07:22,320 --> 01:07:24,200
If the conversation shifts to a new topic,

2005
01:07:24,200 --> 01:07:25,920
the agent swaps the expertise.

2006
01:07:25,920 --> 01:07:28,520
This shift ripples through every part of the system.

2007
01:07:28,520 --> 01:07:30,440
The efficiency gain is the most obvious part.

2008
01:07:30,440 --> 01:07:31,640
We've seen the data.

2009
01:07:31,640 --> 01:07:33,880
Lean context windows lead to faster inference.

2010
01:07:33,880 --> 01:07:36,520
Using fewer tokens per turn lowers your operational costs

2011
01:07:36,520 --> 01:07:39,240
and fewer reasoning loops mean less latency for the user.

2012
01:07:39,240 --> 01:07:42,160
But this isn't just about speed, it's about sustainability.

2013
01:07:42,160 --> 01:07:45,200
A monolithic agent gets heavier as your organization grows.

2014
01:07:45,200 --> 01:07:46,720
You add more policies and procedures,

2015
01:07:46,720 --> 01:07:49,880
the instruction file balloons, and performance starts to degrade.

2016
01:07:49,880 --> 01:07:53,120
Eventually, the cost per conversation rises until you hit a wall.

2017
01:07:53,120 --> 01:07:54,800
A modular agent grows differently.

2018
01:07:54,800 --> 01:07:57,280
You can add new skills, but the baseline stays lean.

2019
01:07:57,280 --> 01:08:00,600
The system scales because you aren't forcing every single conversation

2020
01:08:00,600 --> 01:08:03,360
to carry the weight of every piece of knowledge you own.

2021
01:08:03,360 --> 01:08:06,200
The governance gain is quieter, but it's just as important.

2022
01:08:06,200 --> 01:08:09,760
When your rules are modular, ownership is distributed across the company.

2023
01:08:09,760 --> 01:08:13,080
A skill for leave eligibility stays under HR's control.

2024
01:08:13,080 --> 01:08:14,760
An approval workflow belongs to finance.

2025
01:08:14,760 --> 01:08:18,280
Each team manages its own expertise, so updates happen independently.

2026
01:08:18,280 --> 01:08:20,720
When HR changes a rule, they update their skill

2027
01:08:20,720 --> 01:08:23,040
without worrying about breaking a finance workflow.

2028
01:08:23,040 --> 01:08:24,760
This makes everything auditable.

2029
01:08:24,760 --> 01:08:29,160
You can trace any decision back to a specific version of a skill at a specific time.

2030
01:08:29,160 --> 01:08:31,800
If you need to see every skill that touches customer data,

2031
01:08:31,800 --> 01:08:33,280
you can find them in seconds.

2032
01:08:33,280 --> 01:08:35,800
Modular ownership creates modular accountability.

2033
01:08:35,800 --> 01:08:37,800
Reliability ties all of this together.

2034
01:08:37,800 --> 01:08:40,960
When an agent loads targeted guidance at the right moment,

2035
01:08:40,960 --> 01:08:42,120
it makes better decisions.

2036
01:08:42,120 --> 01:08:43,840
The signal to noise ratio improves,

2037
01:08:43,840 --> 01:08:46,440
because the model isn't drowning in irrelevant data.

2038
01:08:46,440 --> 01:08:49,720
When you measure this, accuracy improves consistently.

2039
01:08:49,720 --> 01:08:51,280
The effect compounds over time.

2040
01:08:51,280 --> 01:08:53,080
Better guidance leads to fewer mistakes

2041
01:08:53,080 --> 01:08:55,080
and fewer turns to reach a resolution.

2042
01:08:55,080 --> 01:08:56,920
The scalability gain is structural.

2043
01:08:56,920 --> 01:08:59,240
A monolithic agent eventually hits a hard limit.

2044
01:08:59,240 --> 01:09:01,760
If you add too many tools or too much knowledge,

2045
01:09:01,760 --> 01:09:03,360
the orchestrator becomes unreliable.

2046
01:09:03,360 --> 01:09:06,360
It loses the path because too many directions are possible at once.

2047
01:09:06,360 --> 01:09:09,160
A modular agent scales because every skill is independent.

2048
01:09:09,160 --> 01:09:12,800
You can run 10 skills or a thousand without the complexity becoming exponential.

2049
01:09:12,800 --> 01:09:17,440
You can handle more users because the token load for each conversation is predictable and lean.

2050
01:09:17,440 --> 01:09:19,240
All of this matters for your AI roadmap,

2051
01:09:19,240 --> 01:09:21,760
because it's the difference between a tactical experiment

2052
01:09:21,760 --> 01:09:23,360
and a sustainable operating model.

2053
01:09:23,360 --> 01:09:26,480
Experiments prove value, but they always hit a wall.

2054
01:09:26,480 --> 01:09:29,920
Operating models scale because they are built for growth from day one.

2055
01:09:29,920 --> 01:09:33,160
A modular skills architecture is how you move from one to the other.

2056
01:09:33,160 --> 01:09:34,280
Common mistakes.

2057
01:09:34,280 --> 01:09:37,440
Teams building skills for the first time usually hit the same walls.

2058
01:09:37,440 --> 01:09:39,080
It's not because the concept is difficult,

2059
01:09:39,080 --> 01:09:40,680
it's because the details matter

2060
01:09:40,680 --> 01:09:43,040
and details are where your intention meets reality.

2061
01:09:43,040 --> 01:09:46,200
The first mistake is treating skill descriptions like documentation.

2062
01:09:46,200 --> 01:09:48,000
Most makers write descriptions for humans.

2063
01:09:48,000 --> 01:09:49,120
They write things like,

2064
01:09:49,120 --> 01:09:52,840
"This skill helps with HR questions about pay, benefits and time off."

2065
01:09:52,840 --> 01:09:54,360
That sounds like a help file.

2066
01:09:54,360 --> 01:09:56,240
It tells a person what the skill does,

2067
01:09:56,240 --> 01:09:57,760
but it does nothing for the orchestrator.

2068
01:09:57,760 --> 01:10:01,240
The orchestrator has to decide if this skill matches what the user wants.

2069
01:10:01,240 --> 01:10:03,040
A vague description creates noise.

2070
01:10:03,040 --> 01:10:04,880
Is this the right skill for leaf questions?

2071
01:10:04,880 --> 01:10:07,480
Maybe. Is it for benefits enrollment? Probably.

2072
01:10:07,480 --> 01:10:10,560
So when a user asks about benefits, the orchestrator gets confused.

2073
01:10:10,560 --> 01:10:12,960
It loads both skills, or it guesses wrong,

2074
01:10:12,960 --> 01:10:14,920
and loads the leaf skill when it shouldn't.

2075
01:10:14,920 --> 01:10:18,520
Your description needs to be a rooting signal, not marketing copy.

2076
01:10:18,520 --> 01:10:20,760
It must be precise, bounded.

2077
01:10:20,760 --> 01:10:24,320
It needs to say exactly when to use it and when to stay away.

2078
01:10:24,320 --> 01:10:26,760
Use for questions about a cruel and blackout periods.

2079
01:10:26,760 --> 01:10:28,960
Do not use for benefits enrollment.

2080
01:10:28,960 --> 01:10:30,680
Now the orchestrator has clarity.

2081
01:10:30,680 --> 01:10:32,320
The second mistake is the mega skill.

2082
01:10:32,320 --> 01:10:35,720
This happens when teams decide all of HR should be one single skill.

2083
01:10:35,720 --> 01:10:37,800
Leave, benefits, pay and policy.

2084
01:10:37,800 --> 01:10:40,000
All bundled into one massive instruction set.

2085
01:10:40,000 --> 01:10:42,040
The skill fires on every HR question,

2086
01:10:42,040 --> 01:10:44,920
then the model has to struggle to find the right subdomain.

2087
01:10:44,920 --> 01:10:46,880
You've created reasoning overhead you didn't need.

2088
01:10:46,880 --> 01:10:50,640
You just moved the monolithic problem from the agent level down to the skill level.

2089
01:10:50,640 --> 01:10:52,360
Mega skills are slow to update.

2090
01:10:52,360 --> 01:10:55,360
Changing one rule about leave might break how the whole skill behaves.

2091
01:10:55,360 --> 01:10:59,240
They are hard to test because you're validating five different domains in one unit.

2092
01:10:59,240 --> 01:11:00,400
The fix is to split them.

2093
01:11:00,400 --> 01:11:03,840
Leave is a skill, benefits is a skill, compensation is a skill.

2094
01:11:03,840 --> 01:11:05,160
Each one loads independently.

2095
01:11:05,160 --> 01:11:06,880
Each one is testable in isolation.

2096
01:11:06,880 --> 01:11:10,240
And each one can be owned by the team that actually knows the data.

2097
01:11:10,240 --> 01:11:12,280
The third mistake is ignoring the governance surface.

2098
01:11:12,280 --> 01:11:15,120
You bundle your scripts, you test them, you push the production,

2099
01:11:15,120 --> 01:11:16,400
but you never review the code.

2100
01:11:16,400 --> 01:11:17,400
You never sandbox it.

2101
01:11:17,400 --> 01:11:18,640
You never look at the logs.

2102
01:11:18,640 --> 01:11:21,840
You just assume the script does what the instructions say it does.

2103
01:11:21,840 --> 01:11:22,680
Then something breaks.

2104
01:11:22,680 --> 01:11:23,520
Data leaks out.

2105
01:11:23,520 --> 01:11:25,040
A credential is exposed.

2106
01:11:25,040 --> 01:11:26,240
Performance drops for no reason.

2107
01:11:26,240 --> 01:11:27,640
You have no audit trail.

2108
01:11:27,640 --> 01:11:29,760
You don't know what the script actually did.

2109
01:11:29,760 --> 01:11:31,120
So you can't trace the incident.

2110
01:11:31,120 --> 01:11:34,480
This happens because teams think scripts are just implementation details.

2111
01:11:34,480 --> 01:11:36,120
They think governance is just overhead.

2112
01:11:36,120 --> 01:11:37,240
They are wrong.

2113
01:11:37,240 --> 01:11:39,720
Scripts are untrusted until you prove otherwise.

2114
01:11:39,720 --> 01:11:41,080
Governance isn't an add-on.

2115
01:11:41,080 --> 01:11:42,160
It's the foundation.

2116
01:11:42,160 --> 01:11:43,120
Review the code.

2117
01:11:43,120 --> 01:11:43,880
Sandbox it.

2118
01:11:43,880 --> 01:11:44,760
Monitor the output.

2119
01:11:44,760 --> 01:11:46,160
If you aren't willing to do that,

2120
01:11:46,160 --> 01:11:48,360
don't bundle scripts.

2121
01:11:48,360 --> 01:11:52,000
The fourth mistake is failing to measure the before and after.

2122
01:11:52,000 --> 01:11:54,520
You roll out skills and just assume they're helping.

2123
01:11:54,520 --> 01:11:56,560
Three months later, you realize you aren't sure.

2124
01:11:56,560 --> 01:11:57,920
Did you actually save tokens?

2125
01:11:57,920 --> 01:11:59,280
Did the latency go down?

2126
01:11:59,280 --> 01:12:01,320
Are the agents actually better at their jobs?

2127
01:12:01,320 --> 01:12:03,720
You don't know because you never set a baseline.

2128
01:12:03,720 --> 01:12:04,960
Measurement takes discipline.

2129
01:12:04,960 --> 01:12:07,240
You have to run the tests before you change anything

2130
01:12:07,240 --> 01:12:08,800
and then run them again after.

2131
01:12:08,800 --> 01:12:11,040
It's the only way to know if the work actually mattered.

2132
01:12:11,040 --> 01:12:13,640
Without numbers, you're just flying on intuition.

2133
01:12:13,640 --> 01:12:14,880
And intuition is expensive.

2134
01:12:14,880 --> 01:12:17,520
The fifth mistake is forgetting that skills are guides,

2135
01:12:17,520 --> 01:12:19,000
not guarantees.

2136
01:12:19,000 --> 01:12:20,920
A skill gives the model expert reasoning.

2137
01:12:20,920 --> 01:12:22,280
But the model is still a model.

2138
01:12:22,280 --> 01:12:23,800
It can still misinterpret the guide.

2139
01:12:23,800 --> 01:12:26,200
It can still hallucinate details that aren't there.

2140
01:12:26,200 --> 01:12:28,840
It can ignore your instructions and just improvise.

2141
01:12:28,840 --> 01:12:30,880
A skill makes a good outcome more likely,

2142
01:12:30,880 --> 01:12:32,280
but it doesn't make it certain.

2143
01:12:32,280 --> 01:12:33,760
Some teams learn this the hard way.

2144
01:12:33,760 --> 01:12:36,240
They deploy a skill to enforce a critical process

2145
01:12:36,240 --> 01:12:38,080
only to find out the model skipped a step.

2146
01:12:38,080 --> 01:12:40,000
The protection wasn't real.

2147
01:12:40,000 --> 01:12:41,440
Skills are guidance layers.

2148
01:12:41,440 --> 01:12:42,600
They are not guardrails.

2149
01:12:42,600 --> 01:12:44,600
If you need to enforce a rule absolutely,

2150
01:12:44,600 --> 01:12:47,280
you need a structural guardrail, a workflow step,

2151
01:12:47,280 --> 01:12:50,760
an approval gate, something deterministic.

2152
01:12:50,760 --> 01:12:53,160
The skill can guide the thinking that leads to the gate,

2153
01:12:53,160 --> 01:12:56,520
but the gate itself has to be part of the architecture.

2154
01:12:56,520 --> 01:12:58,760
The architecture principle, skills aren't just

2155
01:12:58,760 --> 01:13:00,480
a feature you add to an agent.

2156
01:13:00,480 --> 01:13:01,840
They are a structural principle.

2157
01:13:01,840 --> 01:13:02,960
The principle is simple.

2158
01:13:02,960 --> 01:13:05,640
Load only what you need exactly when you need it.

2159
01:13:05,640 --> 01:13:07,320
This applies to your instructions.

2160
01:13:07,320 --> 01:13:08,640
It applies to your knowledge.

2161
01:13:08,640 --> 01:13:10,160
It applies to your tools.

2162
01:13:10,160 --> 01:13:11,840
And to the reasoning itself, the payoff

2163
01:13:11,840 --> 01:13:13,440
is something you can actually measure.

2164
01:13:13,440 --> 01:13:16,400
Faster agents, lower costs, higher reliability.

2165
01:13:16,400 --> 01:13:18,440
Most organizations are still building monoliths.

2166
01:13:18,440 --> 01:13:19,560
That is the status quo.

2167
01:13:19,560 --> 01:13:20,880
But now you know the difference.

2168
01:13:20,880 --> 01:13:22,280
The question is what you do with it.

2169
01:13:22,280 --> 01:13:24,640
If this changed how you think about AI agents,

2170
01:13:24,640 --> 01:13:26,280
follow me, mycopeters on LinkedIn.

2171
01:13:26,280 --> 01:13:28,600
And if you want more of this, leave a review.

2172
01:13:28,600 --> 01:13:30,200
It helps more people find it.