1
00:00:00,000 --> 00:00:03,840
Today's topic is one that almost everyone has heard of, but few can actually explain.

2
00:00:03,840 --> 00:00:05,600
The AI Agent Harness.

3
00:00:05,600 --> 00:00:08,960
You've probably seen it in blog posts or conference talks, but what does it actually

4
00:00:08,960 --> 00:00:09,960
mean?

5
00:00:09,960 --> 00:00:10,960
Here's the thing.

6
00:00:10,960 --> 00:00:14,880
Most people think building an AI agent is about writing the perfect prompt, that one

7
00:00:14,880 --> 00:00:18,160
magical instruction that makes the model do exactly what you want.

8
00:00:18,160 --> 00:00:20,440
But that's like, thinking a car is just an engine.

9
00:00:20,440 --> 00:00:24,580
Sure, the engine is important, but without wheels, a steering wheel breaks and a chassis,

10
00:00:24,580 --> 00:00:25,780
you're not going anywhere.

11
00:00:25,780 --> 00:00:29,880
By the end of this episode, you'll understand what an agent harness actually is.

12
00:00:29,880 --> 00:00:35,140
Why it's the next evolution beyond prompt engineering and how Microsoft AI Foundry brings

13
00:00:35,140 --> 00:00:37,440
this to life for real businesses?

14
00:00:37,440 --> 00:00:39,720
Grab your coffee and let's dive in.

15
00:00:39,720 --> 00:00:41,360
Why prompting isn't enough anymore?

16
00:00:41,360 --> 00:00:42,840
Let's go back to the early days.

17
00:00:42,840 --> 00:00:46,080
Around 2022 to 2024, working with AI was simple.

18
00:00:46,080 --> 00:00:47,760
You wrote a prompt and got an answer.

19
00:00:47,760 --> 00:00:49,040
Need a draft email?

20
00:00:49,040 --> 00:00:50,040
Write a prompt.

21
00:00:50,040 --> 00:00:51,480
Need a summary of an article?

22
00:00:51,480 --> 00:00:54,560
Write a prompt for simple one-shot tasks that worked really well.

23
00:00:54,560 --> 00:00:57,920
You could ask for something, and the model would give you a reasonable response.

24
00:00:57,920 --> 00:00:59,240
But then people got more ambitious.

25
00:00:59,240 --> 00:01:02,240
They started asking these models to do real work.

26
00:01:02,240 --> 00:01:06,960
Research a topic and write a report, analyze a codebase, and fix a bug or handle a multi-step

27
00:01:06,960 --> 00:01:10,040
customer support conversation from start to finish.

28
00:01:10,040 --> 00:01:12,080
And suddenly, one prompt couldn't handle it.

29
00:01:12,080 --> 00:01:13,080
Here's what happened.

30
00:01:13,080 --> 00:01:14,680
The model's context window filled up.

31
00:01:14,680 --> 00:01:18,600
It forgot the earlier instructions, and it started making up facts or giving incomplete

32
00:01:18,600 --> 00:01:19,600
results.

33
00:01:19,600 --> 00:01:22,680
The longer the task went on, the more unreliable the output became.

34
00:01:22,680 --> 00:01:25,000
You've probably experienced this yourself.

35
00:01:25,000 --> 00:01:27,080
You ask an AI to do something complex.

36
00:01:27,080 --> 00:01:30,800
And by the end, it's forgotten what you asked it to do in the first place.

37
00:01:30,800 --> 00:01:35,040
Imagine asking a single person to build an entire house from scratch, all in one sitting,

38
00:01:35,040 --> 00:01:36,760
without ever checking their work.

39
00:01:36,760 --> 00:01:38,560
That's what we were asking these models to do.

40
00:01:38,560 --> 00:01:40,400
The results were exactly what you'd expect.

41
00:01:40,400 --> 00:01:43,240
A half finished house with missing walls and doors that don't open.

42
00:01:43,240 --> 00:01:44,560
The solution wasn't a better prompt.

43
00:01:44,560 --> 00:01:49,320
You could spend hours tweaking the wording, adding more examples, and refining the instructions,

44
00:01:49,320 --> 00:01:52,960
and it would still fail on long complex tasks because the fundamental problem wasn't

45
00:01:52,960 --> 00:01:53,960
the prompt.

46
00:01:53,960 --> 00:01:56,280
It was the lack of a system around it.

47
00:01:56,280 --> 00:01:59,360
But the model needed wasn't better words, but better support.

48
00:01:59,360 --> 00:02:03,120
That's where the evolution from prompt engineering to context engineering began.

49
00:02:03,120 --> 00:02:05,160
The three phases of AI evolution.

50
00:02:05,160 --> 00:02:07,480
So here's how I see the evolution of AI.

51
00:02:07,480 --> 00:02:08,480
Three phases.

52
00:02:08,480 --> 00:02:10,720
Each one shifts the focus for engineers.

53
00:02:10,720 --> 00:02:14,640
Phase one was prompt engineering that ran from about 2022 to 2024.

54
00:02:14,640 --> 00:02:16,080
The big question was simple.

55
00:02:16,080 --> 00:02:17,840
What should I say to the model?

56
00:02:17,840 --> 00:02:21,040
You'd write one instruction, get one answer, and that was it.

57
00:02:21,040 --> 00:02:23,000
The whole craft was about wording and tone.

58
00:02:23,000 --> 00:02:24,920
For basic tasks, that was enough.

59
00:02:24,920 --> 00:02:26,240
Phase two came next.

60
00:02:26,240 --> 00:02:31,280
In context engineering from 2024 through 2025, the question changed from what should I say

61
00:02:31,280 --> 00:02:33,320
to what should the model see.

62
00:02:33,320 --> 00:02:36,760
Engineers found that what you feed the model matters more than how you ask.

63
00:02:36,760 --> 00:02:40,960
So they built systems like tool calling, retrieval augmented generation, and the model context

64
00:02:40,960 --> 00:02:42,680
protocol, MCP.

65
00:02:42,680 --> 00:02:45,480
These let the model grab the right information on the fly.

66
00:02:45,480 --> 00:02:49,280
Instead of stuffing everything into one prompt, the model could fetch what it needed when

67
00:02:49,280 --> 00:02:50,280
it needed it.

68
00:02:50,280 --> 00:02:51,760
Phase three is where we are now.

69
00:02:51,760 --> 00:02:53,440
Harness engineering.

70
00:02:53,440 --> 00:02:56,200
Starting around 2026, the question is, what system do

71
00:02:56,200 --> 00:02:58,000
I need to build around the model?

72
00:02:58,000 --> 00:02:59,320
This is where it gets interesting.

73
00:02:59,320 --> 00:03:03,160
Because reliability, memory, orchestration, and guardrails don't come from the model itself,

74
00:03:03,160 --> 00:03:05,480
they come from the system you build around it.

75
00:03:05,480 --> 00:03:07,320
Here it is, the simplest way to think about it.

76
00:03:07,320 --> 00:03:09,240
An agent equals a model plus a harness.

77
00:03:09,240 --> 00:03:10,440
The harness is everything else.

78
00:03:10,440 --> 00:03:14,000
The model does the thinking, the harness does the rest, and that harness is the secret source

79
00:03:14,000 --> 00:03:15,720
that makes the whole thing work.

80
00:03:15,720 --> 00:03:16,720
Think of it like this.

81
00:03:16,720 --> 00:03:18,320
The model is a skilled worker.

82
00:03:18,320 --> 00:03:19,600
Really talented but limited.

83
00:03:19,600 --> 00:03:24,040
The harness is the workshop, the tools, the checklist, the supervisor, and the brakes.

84
00:03:24,040 --> 00:03:26,160
Without the workshop, the worker has nowhere to work.

85
00:03:26,160 --> 00:03:28,360
Without the tools, they can't build anything.

86
00:03:28,360 --> 00:03:30,000
Without the checklist, they forget steps.

87
00:03:30,000 --> 00:03:31,680
Without the supervisor, they make mistakes.

88
00:03:31,680 --> 00:03:33,200
And without brakes, they burn out.

89
00:03:33,200 --> 00:03:34,840
The harness provides all that structure.

90
00:03:34,840 --> 00:03:36,400
So what exactly is inside this harness?

91
00:03:36,400 --> 00:03:37,640
Let's open it up.

92
00:03:37,640 --> 00:03:39,440
What a harness actually contains.

93
00:03:39,440 --> 00:03:41,120
So what goes into a harness?

94
00:03:41,120 --> 00:03:42,920
Let's walk through the pieces one by one.

95
00:03:42,920 --> 00:03:44,120
First, the agent loop.

96
00:03:44,120 --> 00:03:46,280
This is the cycle that keeps the agent going.

97
00:03:46,280 --> 00:03:48,320
The model thinks then decides to call a tool.

98
00:03:48,320 --> 00:03:50,040
The tool runs and sends back a result.

99
00:03:50,040 --> 00:03:53,160
The model looks at that, thinks again, and maybe calls another tool.

100
00:03:53,160 --> 00:03:56,480
The loop keeps going until the agent reaches its goal.

101
00:03:56,480 --> 00:04:00,120
Without this loop, you just get a single question and a single answer.

102
00:04:00,120 --> 00:04:03,880
With it, you have a system that works through problems step by step.

103
00:04:03,880 --> 00:04:05,560
Next, context management.

104
00:04:05,560 --> 00:04:08,480
This is critical because models have a limited attention span.

105
00:04:08,480 --> 00:04:11,360
Their context window can only hold so much before overflow.

106
00:04:11,360 --> 00:04:12,560
The harness steps in.

107
00:04:12,560 --> 00:04:17,680
It compresses the conversation, summarizes older messages, and prioritizes what's important.

108
00:04:17,680 --> 00:04:19,920
Think of it like a filing system for the conversation.

109
00:04:19,920 --> 00:04:22,520
The harness organizes it so the model only sees what's relevant.

110
00:04:22,520 --> 00:04:25,960
Without this, long sessions fall apart as the model loses track.

111
00:04:25,960 --> 00:04:27,400
Then there are tools and skills.

112
00:04:27,400 --> 00:04:28,600
These are the hands of the agent.

113
00:04:28,600 --> 00:04:30,640
A model by itself can only generate text.

114
00:04:30,640 --> 00:04:33,880
It can't read a file, browse the web, run code, or query a database.

115
00:04:33,880 --> 00:04:35,880
The harness gives it those capabilities.

116
00:04:35,880 --> 00:04:42,160
File access, web browsing, code execution, database queries, all tools the harness provides.

117
00:04:42,160 --> 00:04:45,200
The model decides when to use them and the harness makes them work.

118
00:04:45,200 --> 00:04:47,360
Memory and session persistence is another big piece.

119
00:04:47,360 --> 00:04:49,320
The harness remembers past sessions.

120
00:04:49,320 --> 00:04:52,280
So when you come back to an agent tomorrow, it doesn't start from zero.

121
00:04:52,280 --> 00:04:55,760
It remembers your preferences, your project context, the decisions you made.

122
00:04:55,760 --> 00:04:58,960
The agent learns over time because the harness keeps that history.

123
00:04:58,960 --> 00:05:00,400
Guardrails are the safety layer.

124
00:05:00,400 --> 00:05:02,240
Limits you set on the agent's behavior.

125
00:05:02,240 --> 00:05:04,200
Think of it as the agent's rule book.

126
00:05:04,200 --> 00:05:05,720
Maximum steps before it stops.

127
00:05:05,720 --> 00:05:07,520
Actions it's not allowed to take.

128
00:05:07,520 --> 00:05:11,800
Human approval gates for high-risk moves like writing files or sending emails.

129
00:05:11,800 --> 00:05:15,720
Without guardrails, an agent could run forever or do something you didn't intend.

130
00:05:15,720 --> 00:05:17,160
The harness keeps it in check.

131
00:05:17,160 --> 00:05:18,480
Finally, orchestration.

132
00:05:18,480 --> 00:05:22,080
The harness can spin up specialised sub-agents for different tasks.

133
00:05:22,080 --> 00:05:25,920
One handles research, another handles writing, a third handles verification.

134
00:05:25,920 --> 00:05:29,000
They work together like a team of specialists on a construction site.

135
00:05:29,000 --> 00:05:30,720
Each does what it does best.

136
00:05:30,720 --> 00:05:33,320
And the harness makes sure they don't get in each other's way.

137
00:05:33,320 --> 00:05:35,680
Now these components aren't just theory.

138
00:05:35,680 --> 00:05:39,120
Microsoft has built a full platform that puts all of this into practice.

139
00:05:39,120 --> 00:05:42,520
Enter Microsoft AI Foundry, the Enterprise harness.

140
00:05:42,520 --> 00:05:44,880
That platform is Microsoft AI Foundry.

141
00:05:44,880 --> 00:05:49,120
Think of it as a complete workshop for building and running AI agents at scale.

142
00:05:49,120 --> 00:05:52,440
When I say scale, I mean real scale, not a lab experiment.

143
00:05:52,440 --> 00:05:54,800
Foundry serves over 70,000 customers today.

144
00:05:54,800 --> 00:05:59,160
And last quarter alone, it processed 100 trillion tokens.

145
00:05:59,160 --> 00:06:01,320
That's 100 trillion, not a typo.

146
00:06:01,320 --> 00:06:04,280
Every day, it powers 2 billion Enterprise search queries.

147
00:06:04,280 --> 00:06:07,720
This is a production platform handling some of the biggest workloads in the world.

148
00:06:07,720 --> 00:06:11,160
What Foundry actually does is give you the harness as a managed service.

149
00:06:11,160 --> 00:06:15,320
You don't build context management, memory, guardrails and coordination from scratch because

150
00:06:15,320 --> 00:06:16,400
it's all baked in.

151
00:06:16,400 --> 00:06:19,440
You bring your agent and Foundry provides the workshop around it.

152
00:06:19,440 --> 00:06:20,760
Now let's talk about identity.

153
00:06:20,760 --> 00:06:21,840
This is a big deal.

154
00:06:21,840 --> 00:06:25,000
Every agent in Foundry gets its own Entra agent ID.

155
00:06:25,000 --> 00:06:26,760
Basically a digital identity.

156
00:06:26,760 --> 00:06:29,720
Think of it like an employee badge that the reception desk issues.

157
00:06:29,720 --> 00:06:34,120
The agent can log into services, access data and act on behalf of the organization just like

158
00:06:34,120 --> 00:06:35,360
a human employee.

159
00:06:35,360 --> 00:06:38,880
And because every action is tied to that identity, it's fully auditable.

160
00:06:38,880 --> 00:06:39,880
Why does that matter?

161
00:06:39,880 --> 00:06:42,000
In an enterprise, you need to know who did what.

162
00:06:42,000 --> 00:06:44,040
Every action is logged intracable.

163
00:06:44,040 --> 00:06:49,520
It connects to over 1 400 Enterprise data sources from SharePoint and Dynamics 365 to Salesforce

164
00:06:49,520 --> 00:06:50,880
and Custom databases.

165
00:06:50,880 --> 00:06:53,640
If your company uses it, Foundry probably has a connector.

166
00:06:53,640 --> 00:06:55,720
So the agent isn't guessing or making things up.

167
00:06:55,720 --> 00:06:59,800
It pulls real data from real systems, which means the answers are grounded in your actual

168
00:06:59,800 --> 00:07:00,800
business data.

169
00:07:00,800 --> 00:07:01,800
Then there's memory.

170
00:07:01,800 --> 00:07:03,840
Foundry includes built-in memory types.

171
00:07:03,840 --> 00:07:06,160
Procedural memory helps the agent learn from past tasks.

172
00:07:06,160 --> 00:07:08,840
It gets better over time, like on the job training.

173
00:07:08,840 --> 00:07:12,080
Session persistence means it remembers previous conversations, so it's not starting from

174
00:07:12,080 --> 00:07:13,440
zero every time you talk to it.

175
00:07:13,440 --> 00:07:17,520
It makes interactions feel continuous and intelligent, so how does this actually work?

176
00:07:17,520 --> 00:07:21,320
Let's break down the components Microsoft provides for building custom harnesses.

177
00:07:21,320 --> 00:07:24,280
The agent framework, building custom harnesses.

178
00:07:24,280 --> 00:07:26,520
So how do you actually build a harness?

179
00:07:26,520 --> 00:07:30,680
Microsoft provides a dedicated SDK called the Microsoft agent framework, which you might

180
00:07:30,680 --> 00:07:33,160
know by its earlier name, semantic kernel.

181
00:07:33,160 --> 00:07:34,840
It's been rebuilt, but the idea is the same.

182
00:07:34,840 --> 00:07:38,440
It lets developers build custom agent harnesses in Python and CSAT.

183
00:07:38,440 --> 00:07:39,960
The framework has three layers.

184
00:07:39,960 --> 00:07:44,280
But the bottom is the agent loop, the core reasoning cycle where the model thinks, calls tools,

185
00:07:44,280 --> 00:07:45,920
gets results, and thinks again.

186
00:07:45,920 --> 00:07:49,840
On top of that, sit workflows, coordination patterns that manage multiple agents.

187
00:07:49,840 --> 00:07:53,960
And wrapping everything is the harness layer itself, the shell that holds it all together.

188
00:07:53,960 --> 00:07:57,240
Microsoft ships five built-in coordination patterns.

189
00:07:57,240 --> 00:08:00,640
Sequential runs agents one after another in a defined order.

190
00:08:00,640 --> 00:08:04,320
Concerned runs them all in parallel, handoff passes control from one agent to another,

191
00:08:04,320 --> 00:08:05,600
based on what's needed.

192
00:08:05,600 --> 00:08:08,920
Group chat lets multiple agents talk in a shared conversation.

193
00:08:08,920 --> 00:08:13,040
And then there's magentech, a manager pattern from Microsoft Research here, a supervisor

194
00:08:13,040 --> 00:08:17,760
agent generates a plan, and then delegates work to specialize sub agents like a project

195
00:08:17,760 --> 00:08:20,120
manager assigning tasks to team members.

196
00:08:20,120 --> 00:08:22,520
Each pattern handles a different coordination problem.

197
00:08:22,520 --> 00:08:25,000
Here's the best part, the framework is extensible.

198
00:08:25,000 --> 00:08:29,920
You can plug in any model you want from OpenAI, Anthropic, and Google Gemini to Amazon

199
00:08:29,920 --> 00:08:32,640
Bedrock and even local models on Olamma.

200
00:08:32,640 --> 00:08:37,600
Same for tools, you connect via OpenAPI, the model context protocol or direct code.

201
00:08:37,600 --> 00:08:41,240
Some people worry that using Microsoft's framework locks you into their ecosystem.

202
00:08:41,240 --> 00:08:42,240
That's not true.

203
00:08:42,240 --> 00:08:44,760
The framework works with whatever stack you're already using.

204
00:08:44,760 --> 00:08:49,120
The harness layer adds capabilities like file access, code execution, planning, middleware,

205
00:08:49,120 --> 00:08:50,120
and telemetry.

206
00:08:50,120 --> 00:08:53,360
So building an agent with this framework means building a complete system that reads files,

207
00:08:53,360 --> 00:08:56,920
writes code, plans, multi-step tasks, and logs everything for debugging.

208
00:08:56,920 --> 00:08:58,600
It's more than just a chat interface.

209
00:08:58,600 --> 00:08:59,920
Here's what that looks like.

210
00:08:59,920 --> 00:09:02,200
Imagine a coding agent built on this framework.

211
00:09:02,200 --> 00:09:06,040
You give it a task, say, add a new feature to an existing code base.

212
00:09:06,040 --> 00:09:09,200
The agent starts by browsing documentation to understand the API.

213
00:09:09,200 --> 00:09:11,920
Then it writes the code, then it runs tests to make sure nothing broke.

214
00:09:11,920 --> 00:09:15,840
If a test fails, it reads the error, fixes the bug, and runs the tests again, all within

215
00:09:15,840 --> 00:09:20,480
the same harness, looping through the think-act check cycle until the job is done.

216
00:09:20,480 --> 00:09:22,840
Without the harness, you'd manually feed each step.

217
00:09:22,840 --> 00:09:25,560
With it, the agent handles the whole workflow on its own.

218
00:09:25,560 --> 00:09:28,840
One popular example of a harness in action is the claw pattern.

219
00:09:28,840 --> 00:09:31,800
Microsoft has made it available on Foundry, and it's a great demonstration of how these

220
00:09:31,800 --> 00:09:33,480
pieces come together.

221
00:09:33,480 --> 00:09:35,480
They're all digital style agents and Hermes.

222
00:09:35,480 --> 00:09:37,160
Let's talk about claw style agents.

223
00:09:37,160 --> 00:09:41,440
The name comes from open source projects like OpenClaw, these agents live on your machine,

224
00:09:41,440 --> 00:09:44,600
wait for instructions, and then run long tasks on their own.

225
00:09:44,600 --> 00:09:46,960
They're always there ready to jump in when you need them.

226
00:09:46,960 --> 00:09:50,520
Think of them as a personal assistant that never takes a day off.

227
00:09:50,520 --> 00:09:52,040
Microsoft's version is called Hermes.

228
00:09:52,040 --> 00:09:57,120
It runs inside Foundry's hosted agent sandbox and comes with some impressive abilities.

229
00:09:57,120 --> 00:10:00,160
Hermes has its own file system, its own memory, its own set of tools.

230
00:10:00,160 --> 00:10:04,800
It can set up its own maintenance routines, cleaning up old files, organizing its workspace.

231
00:10:04,800 --> 00:10:08,280
And if it realizes it needs a new skill it doesn't have, it can build one on the fly.

232
00:10:08,280 --> 00:10:10,200
Now here's the thing about claw agents.

233
00:10:10,200 --> 00:10:13,040
Cloud architects call them pets, not cattle.

234
00:10:13,040 --> 00:10:14,040
Pets are unique.

235
00:10:14,040 --> 00:10:16,520
You name them, you care for them, you can't easily replace them.

236
00:10:16,520 --> 00:10:19,960
Cattle are interchangeable, so if one goes down, you spin up another.

237
00:10:19,960 --> 00:10:21,480
Claw agents are definitely pets.

238
00:10:21,480 --> 00:10:24,520
Each one has its own state, history, and learned behaviors.

239
00:10:24,520 --> 00:10:28,680
That uniqueness makes them powerful, but it also makes recovery and scaling harder.

240
00:10:28,680 --> 00:10:32,120
If something goes wrong, you can't just throw away the instance and start over because

241
00:10:32,120 --> 00:10:34,280
you'd lose all that accumulated knowledge.

242
00:10:34,280 --> 00:10:36,280
Microsoft's solution is something called routines.

243
00:10:36,280 --> 00:10:40,320
These are scheduled tasks that let the agent wake up, do maintenance work, and then go

244
00:10:40,320 --> 00:10:41,320
back to sleep.

245
00:10:41,320 --> 00:10:44,840
Instead of keeping the sandbox running 24/7, which costs money.

246
00:10:44,840 --> 00:10:47,680
The agent shuts down after a period of inactivity.

247
00:10:47,680 --> 00:10:51,980
The routine system wakes it up when it's time to do something, like clean up old skills,

248
00:10:51,980 --> 00:10:54,160
run a backup, or check for new tasks.

249
00:10:54,160 --> 00:10:55,440
Then it goes back to sleep.

250
00:10:55,440 --> 00:10:58,840
This saves cost while keeping the agent available when you need it.

251
00:10:58,840 --> 00:11:00,800
The sandbox isolates each session.

252
00:11:00,800 --> 00:11:05,880
Your agent's files, memory, and configuration all stay contained in its own environment,

253
00:11:05,880 --> 00:11:09,080
and the harness handles managing that state across sessions, so you don't have to worry

254
00:11:09,080 --> 00:11:11,680
about one agent's data leaking into another's.

255
00:11:11,680 --> 00:11:15,720
Building your own harness from scratch is possible, but most businesses don't need to.

256
00:11:15,720 --> 00:11:19,040
Foundry gives you a ready-made platform with all these capabilities built in.

257
00:11:19,040 --> 00:11:22,920
The real question is, what do you want your agent to do?

258
00:11:22,920 --> 00:11:26,680
Building and collaboration where harnesses meet humans, so you've built your agent given

259
00:11:26,680 --> 00:11:28,880
it tools, memory, orchestration, and a sandbox.

260
00:11:28,880 --> 00:11:30,360
Now how do people actually use it?

261
00:11:30,360 --> 00:11:32,360
Here's where Foundry makes things easy.

262
00:11:32,360 --> 00:11:36,400
Once your agent is built and deployed, you can publish it to Microsoft 365 Co-Pilot and

263
00:11:36,400 --> 00:11:40,640
Microsoft Teams with one click, not 10 clicks, not a three-day deployment pipeline, just

264
00:11:40,640 --> 00:11:41,640
one click.

265
00:11:41,640 --> 00:11:44,280
And your agent shows up right where people already work.

266
00:11:44,280 --> 00:11:46,440
In their email, their chat and their documents.

267
00:11:46,440 --> 00:11:48,080
Foundry supports two deployment modes.

268
00:11:48,080 --> 00:11:49,480
First, assisting agents.

269
00:11:49,480 --> 00:11:53,400
These work on your behalf, drafting emails, scheduling meetings, or pulling up documents.

270
00:11:53,400 --> 00:11:58,720
They act as an extension of you, like having a really fast assistant who does what you ask.

271
00:11:58,720 --> 00:12:00,280
Second, autopilot agents.

272
00:12:00,280 --> 00:12:01,560
These are newer and more interesting.

273
00:12:01,560 --> 00:12:04,920
They act on their own behalf with their own email address and their own identity.

274
00:12:04,920 --> 00:12:08,000
They don't need you to prompt them every time they take initiative.

275
00:12:08,000 --> 00:12:09,840
Autopilot agents have a full user account.

276
00:12:09,840 --> 00:12:12,960
You can add them to a Teams group chat just like you'd add a human coworker.

277
00:12:12,960 --> 00:12:16,920
They follow the conversation and jump in when they can help tracking open items, answering

278
00:12:16,920 --> 00:12:18,960
questions about what the team is working on.

279
00:12:18,960 --> 00:12:23,040
One on-boarding new team members by sending them the right documents and introducing them

280
00:12:23,040 --> 00:12:24,040
to the right people.

281
00:12:24,040 --> 00:12:25,320
Here's a concrete example.

282
00:12:25,320 --> 00:12:27,680
Say you have a "Workstream Manager" agent.

283
00:12:27,680 --> 00:12:29,200
You add it to your Teams group chat.

284
00:12:29,200 --> 00:12:32,080
Someone says, "Hey, I need the Q3 report finished by Friday."

285
00:12:32,080 --> 00:12:34,160
The agent notices this is an action item.

286
00:12:34,160 --> 00:12:38,240
It creates a task, assigns it to the right person, and adds a reminder for Thursday to check

287
00:12:38,240 --> 00:12:39,240
on progress.

288
00:12:39,240 --> 00:12:40,240
Nobody told it to do that.

289
00:12:40,240 --> 00:12:42,800
It just saw something that needed doing and did it.

290
00:12:42,800 --> 00:12:45,520
The harness handles all the permissions behind the scenes.

291
00:12:45,520 --> 00:12:48,160
Only authorized users can interact with the agent in a group chat.

292
00:12:48,160 --> 00:12:50,920
The agent sees messages, it's allowed to see and ignores the rest.

293
00:12:50,920 --> 00:12:53,880
It respects the same access controls your human team members do.

294
00:12:53,880 --> 00:12:55,880
Here's why harness engineering really pays off.

295
00:12:55,880 --> 00:12:58,600
The agent stops being a chatbot you open in a browser tab.

296
00:12:58,600 --> 00:13:00,360
It becomes a productive team member.

297
00:13:00,360 --> 00:13:03,160
It lives where you live, works the way your team works.

298
00:13:03,160 --> 00:13:07,800
And it does all of that because the harness, the system around the model, handles identity,

299
00:13:07,800 --> 00:13:09,560
permissions, deployment and integration.

300
00:13:09,560 --> 00:13:12,560
Without the harness, you'd just have another chat window to check.

301
00:13:12,560 --> 00:13:14,240
But any powerful tool needs safeguards.

302
00:13:14,240 --> 00:13:16,880
Let's talk about the guardrails that make all of this safe.

303
00:13:16,880 --> 00:13:19,240
Responsible AI and governance in the harness.

304
00:13:19,240 --> 00:13:20,240
Here's the thing.

305
00:13:20,240 --> 00:13:22,920
Harness engineering isn't just about making agents more capable.

306
00:13:22,920 --> 00:13:25,160
It's about making them safe and trustworthy.

307
00:13:25,160 --> 00:13:29,400
An agent that can send emails, modify files and acting group chats is powerful, but that

308
00:13:29,400 --> 00:13:30,640
power needs controls.

309
00:13:30,640 --> 00:13:33,120
Think of it like giving an employee keys to the building.

310
00:13:33,120 --> 00:13:36,520
You want them to do their job, but you also set rules about what they can access and

311
00:13:36,520 --> 00:13:37,520
when.

312
00:13:37,520 --> 00:13:41,400
Microsoft's approach is built on the responsible AI standard.

313
00:13:41,400 --> 00:13:43,280
Six principles guide everything.

314
00:13:43,280 --> 00:13:49,640
Harness, reliability and safety, privacy and security, inclusiveness, transparency and accountability.

315
00:13:49,640 --> 00:13:51,720
These aren't just nice ideas written on a poster.

316
00:13:51,720 --> 00:13:55,400
They are the actual foundation for how the entire platform is designed from the ground

317
00:13:55,400 --> 00:13:56,400
up.

318
00:13:56,400 --> 00:13:59,000
The harness is where these principles become real code.

319
00:13:59,000 --> 00:14:00,640
Guardrails limit what the agent can do.

320
00:14:00,640 --> 00:14:04,720
Life cycle hooks let you enforce policies at every step of the agent's execution before

321
00:14:04,720 --> 00:14:06,040
the agent runs.

322
00:14:06,040 --> 00:14:09,880
After it runs, before it calls a tool, after it gets a result, you can inject your own

323
00:14:09,880 --> 00:14:13,800
checks at any point.

324
00:14:13,800 --> 00:14:14,800
You can do that.

325
00:14:14,800 --> 00:14:17,520
Want to require a human approval before the agent sends an email?

326
00:14:17,520 --> 00:14:18,520
You can do that too.

327
00:14:18,520 --> 00:14:20,160
Both of those are straightforward to set up.

328
00:14:20,160 --> 00:14:23,960
As you as well architected framework includes dedicated guidance for AI workloads.

329
00:14:23,960 --> 00:14:27,160
It covers everything from data encryption to incident response.

330
00:14:27,160 --> 00:14:30,640
The idea is that building an AI system isn't fundamentally different from building any

331
00:14:30,640 --> 00:14:31,920
other enterprise system.

332
00:14:31,920 --> 00:14:34,920
You still need security, reliability and operational excellence.

333
00:14:34,920 --> 00:14:38,240
The well architected framework gives you a structured way to think about all of that

334
00:14:38,240 --> 00:14:39,440
without getting lost.

335
00:14:39,440 --> 00:14:41,520
Beyond that, there's the cloud adoption framework.

336
00:14:41,520 --> 00:14:44,280
It provides a four-stage process for AI governance.

337
00:14:44,280 --> 00:14:45,800
First identify risks.

338
00:14:45,800 --> 00:14:46,800
What could go wrong?

339
00:14:46,800 --> 00:14:47,800
Second measure impact.

340
00:14:47,800 --> 00:14:48,800
How bad would it be?

341
00:14:48,800 --> 00:14:50,320
Third mitigate with controls.

342
00:14:50,320 --> 00:14:52,160
What can you put in place to prevent it?

343
00:14:52,160 --> 00:14:53,760
Fourth operate with monitoring.

344
00:14:53,760 --> 00:14:55,360
How do you catch problems when they happen?

345
00:14:55,360 --> 00:14:59,000
It's a practical step-by-step approach that turns abstract concerns into concrete actions

346
00:14:59,000 --> 00:15:00,320
you can actually take.

347
00:15:00,320 --> 00:15:03,440
For agent harnesses specifically, this means a few key things.

348
00:15:03,440 --> 00:15:06,640
Maximum step limits so the agent can't run forever.

349
00:15:06,640 --> 00:15:09,400
Two restrictions so it can only access what it's supposed to.

350
00:15:09,400 --> 00:15:13,080
Human approval gates for high-risk actions audit trails so every decision is logged and

351
00:15:13,080 --> 00:15:18,040
reviewable, content safety filters that catch problematic outputs before anyone sees them.

352
00:15:18,040 --> 00:15:20,760
Every piece works together to keep your agent on a short leash.

353
00:15:20,760 --> 00:15:24,440
Microsoft provides built-in content safety and evaluators inside Foundry.

354
00:15:24,440 --> 00:15:27,360
You can test your agent against a set of criteria before you deploy it.

355
00:15:27,360 --> 00:15:28,600
Does it stay on topic?

356
00:15:28,600 --> 00:15:30,200
Does it handle sensitive data properly?

357
00:15:30,200 --> 00:15:31,200
Does it refuse requests?

358
00:15:31,200 --> 00:15:32,480
It shouldn't fulfill?

359
00:15:32,480 --> 00:15:36,600
You can run these evaluations automatically and get a score before you ever publish the

360
00:15:36,600 --> 00:15:37,920
agent to your team.

361
00:15:37,920 --> 00:15:40,880
It's a bit where you catch problems early, not after something goes wrong.

362
00:15:40,880 --> 00:15:42,120
This is the full picture.

363
00:15:42,120 --> 00:15:46,320
From a simple prompt to a production grade safe autonomous agent, the model does the thinking,

364
00:15:46,320 --> 00:15:50,640
the harness does everything else, the coordination, the guardrails, the governance.

365
00:15:50,640 --> 00:15:52,160
Everything that makes it trustworthy.

366
00:15:52,160 --> 00:15:54,240
So let's recap the transformation.

367
00:15:54,240 --> 00:15:55,880
We started with one-shot prompting.

368
00:15:55,880 --> 00:15:57,720
Write a prompt, get an answer.

369
00:15:57,720 --> 00:15:59,160
That worked for simple tasks.

370
00:15:59,160 --> 00:16:03,160
Then we moved to context engineering, manage what the model sees, load the right information

371
00:16:03,160 --> 00:16:04,160
at the right time.

372
00:16:04,160 --> 00:16:05,960
That worked for more complex tasks.

373
00:16:05,960 --> 00:16:09,840
And now we've arrived at full harness engineering, build a complete system around the model with

374
00:16:09,840 --> 00:16:12,880
memory tools, coordination, guardrails and governance.

375
00:16:12,880 --> 00:16:14,040
Here's the key takeaway.

376
00:16:14,040 --> 00:16:16,680
An agent is only as reliable as the system around it.

377
00:16:16,680 --> 00:16:20,160
The model matters absolutely, but the harness is what makes or breaks production.

378
00:16:20,160 --> 00:16:24,120
You can have the best model in the world, and without a good harness, it will fail on real

379
00:16:24,120 --> 00:16:25,120
tasks.

380
00:16:25,120 --> 00:16:27,640
And you can have a decent model with an excellent harness.

381
00:16:27,640 --> 00:16:30,280
And it will outperform expectations every time.

382
00:16:30,280 --> 00:16:31,680
Here's your challenge.

383
00:16:31,680 --> 00:16:34,680
In your next AI project, think beyond the prompt.

384
00:16:34,680 --> 00:16:36,440
Don't ask what should I say to the model?

385
00:16:36,440 --> 00:16:38,960
Ask what system do I need to build around it?

386
00:16:38,960 --> 00:16:42,000
That shift in thinking is what separates demos from production systems.

387
00:16:42,000 --> 00:16:45,920
In the next episode, we'll dive deeper into one harness component, memory, how agents learn

388
00:16:45,920 --> 00:16:50,880
over time, how they remember what they've done, and how that changes what they can accomplish.

389
00:16:50,880 --> 00:16:51,840
Subscribe so you don't miss it.