1 00:00:00,040 --> 00:00:03,359 Welcome to the M365 FM podcast, where we explore the people, 2 00:00:03,359 --> 00:00:06,400 ideas and technologies shaping the future of work. 3 00:00:06,400 --> 00:00:08,759 Why one AI agent hits a wall. 4 00:00:08,759 --> 00:00:12,119 Multi-Agent sounds like a swarm of robots passing work around 5 00:00:12,119 --> 00:00:13,119 inside Copilot Studio. 6 00:00:13,119 --> 00:00:14,439 It's closer to one agent playing 7 00:00:14,439 --> 00:00:15,519 manager and handing jobs to a 8 00:00:15,519 --> 00:00:16,199 few specialists. 9 00:00:16,199 --> 00:00:19,359 You don't need to be a developer because Copilot Studio builds it 10 00:00:19,359 --> 00:00:21,079 from plain descriptions. 11 00:00:21,079 --> 00:00:22,440 But adding more agents can make 12 00:00:22,440 --> 00:00:23,839 things slower, pricier, and less 13 00:00:23,839 --> 00:00:24,519 accurate. 14 00:00:24,519 --> 00:00:27,039 By the end, you'll know when one agent is enough. 15 00:00:27,039 --> 00:00:27,879 Start with the picture. 16 00:00:27,879 --> 00:00:30,440 Almost everyone has one brilliant assistant that knows 17 00:00:30,440 --> 00:00:33,799 your HR policy answers it tickets, checks, pricing and 18 00:00:33,799 --> 00:00:36,119 books a meeting all from the same chat box. 19 00:00:36,119 --> 00:00:36,840 That's the dream. 20 00:00:36,840 --> 00:00:38,439 And that's also where the trouble starts. 21 00:00:38,439 --> 00:00:40,420 Here's what actually happens when you build it that way you 22 00:00:40,420 --> 00:00:43,679 keep piling jobs onto one agent and every job arrives with tools 23 00:00:43,679 --> 00:00:45,079 and instructions attached. 24 00:00:45,079 --> 00:00:47,500 The agent now has to pick the right action out of a growing 25 00:00:47,500 --> 00:00:50,600 pile, and the picking gets less reliable as the pile grows, 26 00:00:50,600 --> 00:00:51,960 Microsoft puts a number on that. 27 00:00:51,960 --> 00:00:54,640 Its own guidance points to where routing starts to wobble. 28 00:00:54,640 --> 00:00:57,179 Once a single agent is choosing between more than about thirty 29 00:00:57,179 --> 00:01:00,210 to forty actions, counting tools, topics and other agents. 30 00:01:00,210 --> 00:01:01,850 The answers get less dependable. 31 00:01:01,850 --> 00:01:02,570 Past that line. 32 00:01:02,570 --> 00:01:04,689 The agent picks wrong more often, and a wrong pick is hard 33 00:01:04,689 --> 00:01:05,969 to spot until someone complains. 34 00:01:05,969 --> 00:01:08,010 There's one mistake that shows up in nearly every failed 35 00:01:08,010 --> 00:01:11,129 rollout, and I'll get to it once you've seen how the pieces fit, 36 00:01:11,129 --> 00:01:13,730 hold on to that, because it's the part nobody demos. 37 00:01:13,730 --> 00:01:16,730 So the real fix is a smaller team with a clear manager, and 38 00:01:16,730 --> 00:01:18,930 that idea runs through everything else here. 39 00:01:18,930 --> 00:01:21,730 One agent stays at the front door while a few others handle 40 00:01:21,730 --> 00:01:23,250 narrow jobs and report back. 41 00:01:23,250 --> 00:01:24,209 That's what people mean by 42 00:01:24,209 --> 00:01:25,769 multi-agent, and it's a much 43 00:01:25,769 --> 00:01:26,890 calmer thing than the name 44 00:01:26,890 --> 00:01:27,730 suggests. 45 00:01:27,730 --> 00:01:29,969 This matters more now than it did a year ago. 46 00:01:29,969 --> 00:01:32,450 Multi-Agent orchestration arrived in Copilot Studio at 47 00:01:32,450 --> 00:01:35,489 Microsoft Build in May twenty twenty five, and Microsoft has 48 00:01:35,489 --> 00:01:38,329 since moved the multi-agent side toward general availability. 49 00:01:38,329 --> 00:01:39,730 It isn't a demo trick anymore. 50 00:01:39,730 --> 00:01:42,530 It's something your team will be asked to design, and getting the 51 00:01:42,530 --> 00:01:45,569 first decision right saves a painful rebuild later. 52 00:01:45,569 --> 00:01:48,129 It also helps to name why the single big agent feels so 53 00:01:48,129 --> 00:01:49,329 appealing in the first place. 54 00:01:49,329 --> 00:01:51,890 It's a single thing to build, test and explain to your boss, 55 00:01:51,890 --> 00:01:54,849 so splitting it looks like extra work for no visible gain. 56 00:01:54,849 --> 00:01:58,030 That instinct is right more often than the hype suggests. 57 00:01:58,030 --> 00:01:59,909 What Copilot Studio actually is. 58 00:01:59,909 --> 00:02:03,129 Copilot studio is Microsoft's low code studio for building AI 59 00:02:03,129 --> 00:02:04,230 agents and workflows. 60 00:02:04,230 --> 00:02:07,269 Low code means you spend your time describing the job you want 61 00:02:07,269 --> 00:02:09,830 done, instead of writing programming code to make it 62 00:02:09,830 --> 00:02:12,469 happen, you point it at your data and you publish it where 63 00:02:12,469 --> 00:02:13,629 people already work. 64 00:02:13,629 --> 00:02:14,949 An agent is the piece that talks 65 00:02:14,949 --> 00:02:15,990 to a person and gets things 66 00:02:15,990 --> 00:02:16,550 done. 67 00:02:16,550 --> 00:02:18,270 Three things make it run instructions. 68 00:02:18,270 --> 00:02:19,590 Tell it how to behave. 69 00:02:19,590 --> 00:02:20,430 Knowledge sources. 70 00:02:20,430 --> 00:02:23,349 Give it facts to draw on and tools let it take action. 71 00:02:23,349 --> 00:02:24,469 In other systems. 72 00:02:24,469 --> 00:02:25,990 An internal HR helper is the 73 00:02:25,990 --> 00:02:27,110 classic example because it 74 00:02:27,110 --> 00:02:28,669 answers policy questions, looks 75 00:02:28,669 --> 00:02:30,030 up a leave balance, and files a 76 00:02:30,030 --> 00:02:31,349 request all inside one chat 77 00:02:31,349 --> 00:02:31,990 window. 78 00:02:31,990 --> 00:02:34,229 You can build workflows in the same place. 79 00:02:34,229 --> 00:02:35,590 A workflow is an automation. 80 00:02:35,590 --> 00:02:39,150 You assemble step by step and each step can reason and act. 81 00:02:39,150 --> 00:02:41,990 Plenty of business processes are boring and repeatable, so a step 82 00:02:41,990 --> 00:02:45,189 can pull data, run a check or hand a decision to a colleague, 83 00:02:45,189 --> 00:02:46,349 and then continue. 84 00:02:46,349 --> 00:02:47,310 That matters later. 85 00:02:47,310 --> 00:02:48,870 Because a workflow can call an 86 00:02:48,870 --> 00:02:50,750 agent and an agent can call a 87 00:02:50,750 --> 00:02:51,509 workflow. 88 00:02:51,509 --> 00:02:53,710 The channels are the part most people underestimate. 89 00:02:53,710 --> 00:02:56,039 The same agent can show up in Microsoft Teams, inside 90 00:02:56,039 --> 00:02:58,639 Microsoft Copilot on a website or in a mobile app. 91 00:02:58,639 --> 00:03:00,479 It can also work in more than one language. 92 00:03:00,479 --> 00:03:01,860 So you build it once and it 93 00:03:01,860 --> 00:03:03,039 meets people where they already 94 00:03:03,039 --> 00:03:03,520 are. 95 00:03:03,520 --> 00:03:05,439 It's worth being clear about one common mix up. 96 00:03:05,439 --> 00:03:07,960 Copilot studio is where you build and manage agents, and 97 00:03:07,960 --> 00:03:10,919 Microsoft Copilot is one of the places those agents can show up. 98 00:03:10,919 --> 00:03:12,879 They're related, but they aren't the same product. 99 00:03:12,879 --> 00:03:15,680 One more layer matters, and it's the layer that trips people up. 100 00:03:15,680 --> 00:03:18,520 Whatever you build runs on an engine, and Microsoft calls that 101 00:03:18,520 --> 00:03:19,599 engine a harness. 102 00:03:19,599 --> 00:03:22,439 The harness decides how deeply the agent can reason, how 103 00:03:22,439 --> 00:03:25,599 complex a task it can take on, and how your usage gets built. 104 00:03:25,599 --> 00:03:28,439 You don't need to memorize the list of harnesses, you just need 105 00:03:28,439 --> 00:03:31,280 to know that the engine changes both capability and cost. 106 00:03:31,280 --> 00:03:33,520 So it's a real decision rather than a checkbox. 107 00:03:33,520 --> 00:03:35,199 And there's a trap built into 108 00:03:35,199 --> 00:03:36,639 low code that nobody warns you 109 00:03:36,639 --> 00:03:37,039 about. 110 00:03:37,039 --> 00:03:37,960 Because it's low code. 111 00:03:37,960 --> 00:03:40,319 You can get a working agent going in an afternoon. 112 00:03:40,319 --> 00:03:41,520 The same ease means you can 113 00:03:41,520 --> 00:03:42,439 build something that looks 114 00:03:42,439 --> 00:03:44,199 finished and behaves badly, and 115 00:03:44,199 --> 00:03:45,719 that gap is where most of the 116 00:03:45,719 --> 00:03:47,439 pain lives. 117 00:03:47,439 --> 00:03:50,919 What makes an agent multiagent multi-agent orchestration means 118 00:03:50,919 --> 00:03:52,759 one agent calls other agents. 119 00:03:52,759 --> 00:03:54,210 That's the whole definition. 120 00:03:54,210 --> 00:03:57,330 And it's worth saying plainly because the name makes it sound 121 00:03:57,330 --> 00:03:58,330 much bigger than it is. 122 00:03:58,330 --> 00:03:59,490 There's no separate product you 123 00:03:59,490 --> 00:04:00,789 buy and no new platform to 124 00:04:00,789 --> 00:04:01,090 learn. 125 00:04:01,090 --> 00:04:03,449 You just let one agent hand work to others. 126 00:04:03,449 --> 00:04:06,289 Microsoft's own framing is that breaking a problem into 127 00:04:06,289 --> 00:04:08,889 specialized agents makes an application more modular, 128 00:04:08,889 --> 00:04:10,610 scalable, and manageable. 129 00:04:10,610 --> 00:04:14,050 In plain terms, each agent gets a smaller job, so each one is 130 00:04:14,050 --> 00:04:17,370 easier to build well and easier to fix when it misbehaves. 131 00:04:17,370 --> 00:04:20,250 Two flavors exist here, and the names are the confusing part. 132 00:04:20,250 --> 00:04:22,649 Child agents live inside one agent, and they're small and 133 00:04:22,649 --> 00:04:25,529 lightweight, almost like little routines the main agent can run. 134 00:04:25,529 --> 00:04:27,529 Connected agents are full agents 135 00:04:27,529 --> 00:04:29,009 in their own right, with their 136 00:04:29,009 --> 00:04:30,170 own tools and their own 137 00:04:30,170 --> 00:04:31,850 knowledge, and your main agent 138 00:04:31,850 --> 00:04:33,129 reaches out to them across a 139 00:04:33,129 --> 00:04:33,889 boundary. 140 00:04:33,889 --> 00:04:36,329 The difference shows up in how they share context. 141 00:04:36,329 --> 00:04:37,889 A child agent inherits the main 142 00:04:37,889 --> 00:04:39,410 agent's conversation, so passing 143 00:04:39,410 --> 00:04:40,689 information back and forth is 144 00:04:40,689 --> 00:04:41,370 simple. 145 00:04:41,370 --> 00:04:42,870 A connected agent only gets that 146 00:04:42,870 --> 00:04:44,329 history if it's context setting 147 00:04:44,329 --> 00:04:45,810 says so, which is a setting you 148 00:04:45,810 --> 00:04:46,129 check. 149 00:04:46,129 --> 00:04:47,490 Rather than assume. 150 00:04:47,490 --> 00:04:50,050 The mental model that finally makes this click is a manager 151 00:04:50,050 --> 00:04:50,889 and a team. 152 00:04:50,889 --> 00:04:52,899 Your main agent sits at the front door. 153 00:04:52,899 --> 00:04:54,339 It greets the person, figures 154 00:04:54,339 --> 00:04:55,819 out what they need, and routes 155 00:04:55,819 --> 00:04:57,259 the work to whichever specialist 156 00:04:57,259 --> 00:04:58,800 fits those specialists, do 157 00:04:58,800 --> 00:04:59,939 narrow jobs and hand their 158 00:04:59,939 --> 00:05:00,740 findings back. 159 00:05:00,740 --> 00:05:02,579 So nobody has to be brilliant at everything. 160 00:05:02,579 --> 00:05:03,579 A simple example makes that 161 00:05:03,579 --> 00:05:05,259 concrete picture a new employees 162 00:05:05,259 --> 00:05:05,740 first week. 163 00:05:05,740 --> 00:05:08,500 One agent could answer questions about benefits, another could 164 00:05:08,500 --> 00:05:10,939 handle equipment requests, and a third could walk them through 165 00:05:10,939 --> 00:05:12,220 the systems they need. 166 00:05:12,220 --> 00:05:14,899 Each one knows its own area deeply, and the front door agent 167 00:05:14,899 --> 00:05:17,220 decides which one the new hire needs at that moment. 168 00:05:17,220 --> 00:05:19,019 That's why the swamp picture is wrong. 169 00:05:19,019 --> 00:05:20,819 A swarm has no manager, and a 170 00:05:20,819 --> 00:05:22,459 multi-agent setup in Copilot 171 00:05:22,459 --> 00:05:23,300 Studio does. 172 00:05:23,300 --> 00:05:26,060 Almost every problem people run into comes from getting that 173 00:05:26,060 --> 00:05:27,860 manager's job description wrong. 174 00:05:27,860 --> 00:05:28,819 There's one choice here that 175 00:05:28,819 --> 00:05:30,180 decides whether this helps you 176 00:05:30,180 --> 00:05:31,899 or quietly costs you money, and 177 00:05:31,899 --> 00:05:32,899 it's the choice between those 178 00:05:32,899 --> 00:05:33,620 two flavors. 179 00:05:33,620 --> 00:05:34,980 We'll get to it next. 180 00:05:34,980 --> 00:05:36,899 Child agents versus connected agents. 181 00:05:36,899 --> 00:05:39,339 This is the decision that separates a setup that scales 182 00:05:39,339 --> 00:05:40,500 from one that becomes a mess. 183 00:05:40,500 --> 00:05:41,899 So let's take it slowly. 184 00:05:41,899 --> 00:05:44,779 Child agents are the simple option, and simple is usually 185 00:05:44,779 --> 00:05:46,060 the right place to start. 186 00:05:46,060 --> 00:05:49,019 You build them inside the main agent, they share its settings 187 00:05:49,019 --> 00:05:51,439 and you don't have to publish or deploy them separately. 188 00:05:51,439 --> 00:05:54,600 Microsoft's guidance lists the fit clearly use a child agent 189 00:05:54,600 --> 00:05:57,000 for a single clear job, like creating a ticket, checking a 190 00:05:57,000 --> 00:05:58,120 status, or booking a flight. 191 00:05:58,120 --> 00:05:59,759 Use one when a small team owns 192 00:05:59,759 --> 00:06:01,399 the whole solution and nobody 193 00:06:01,399 --> 00:06:02,800 needs to reuse that piece 194 00:06:02,800 --> 00:06:03,759 somewhere else. 195 00:06:03,759 --> 00:06:05,839 Connected agents are the powerful option, and they come 196 00:06:05,839 --> 00:06:06,800 with a price. 197 00:06:06,800 --> 00:06:08,360 A connected agent stands on its own. 198 00:06:08,360 --> 00:06:11,279 It has its own orchestration, tools and knowledge, and the 199 00:06:11,279 --> 00:06:13,720 main agent delegates part of a request to it. 200 00:06:13,720 --> 00:06:16,879 Microsoft's own example is an IT agent that calls a sales agent 201 00:06:16,879 --> 00:06:18,360 to get pricing information. 202 00:06:18,360 --> 00:06:19,800 You reach for this in three situations. 203 00:06:19,800 --> 00:06:21,680 Maybe different teams own different pieces. 204 00:06:21,680 --> 00:06:22,959 Maybe an agent needs its own 205 00:06:22,959 --> 00:06:25,000 settings or release cycle, or 206 00:06:25,000 --> 00:06:26,560 maybe you want to reuse one 207 00:06:26,560 --> 00:06:28,160 capability across several front 208 00:06:28,160 --> 00:06:28,759 doors. 209 00:06:28,759 --> 00:06:31,600 One more thing makes connected agents worth the trouble because 210 00:06:31,600 --> 00:06:32,680 it lives independently. 211 00:06:32,680 --> 00:06:33,740 Several different front door 212 00:06:33,740 --> 00:06:34,759 agents can call the same 213 00:06:34,759 --> 00:06:35,519 specialist. 214 00:06:35,519 --> 00:06:38,120 A sales agent built by one team can serve an onboarding agent 215 00:06:38,120 --> 00:06:40,959 built by another, and when that specialist improves, every front 216 00:06:40,959 --> 00:06:42,240 door improves at the same time. 217 00:06:42,240 --> 00:06:43,600 That reuse is the real prize. 218 00:06:43,600 --> 00:06:46,120 There's a middle path too, and most real solutions use it. 219 00:06:46,120 --> 00:06:49,129 You can mix both kinds in one solution, break a piece out as a 220 00:06:49,129 --> 00:06:50,889 separate agent so people can reach it directly. 221 00:06:50,889 --> 00:06:52,129 Then give that agent its own 222 00:06:52,129 --> 00:06:53,689 child agents for the narrow jobs 223 00:06:53,689 --> 00:06:54,170 inside it. 224 00:06:54,170 --> 00:06:57,250 Microsoft points out that this keeps each layer focused. 225 00:06:57,250 --> 00:06:58,569 The honest trade off is this. 226 00:06:58,569 --> 00:07:00,050 Microsoft reports that connected 227 00:07:00,050 --> 00:07:02,050 agents increase latency because 228 00:07:02,050 --> 00:07:03,490 every hand off adds another 229 00:07:03,490 --> 00:07:04,610 orchestration hop. 230 00:07:04,610 --> 00:07:06,449 The request goes out, the specialist thinks, and the 231 00:07:06,449 --> 00:07:07,290 answer comes back. 232 00:07:07,290 --> 00:07:09,089 And each of those steps takes time. 233 00:07:09,089 --> 00:07:11,930 Microsoft also confirms that connected agents widen the 234 00:07:11,930 --> 00:07:14,350 testing, management and governance surface area, which 235 00:07:14,350 --> 00:07:16,970 in plain words means more moving parts to watch. 236 00:07:16,970 --> 00:07:20,050 So the rule that saves beginners the most pain is this start with 237 00:07:20,050 --> 00:07:23,290 one agent and split only when you can name the boundary. 238 00:07:23,290 --> 00:07:25,930 If you can't say in one sentence why a piece deserves to be its 239 00:07:25,930 --> 00:07:27,449 own agent, it probably doesn't. 240 00:07:27,449 --> 00:07:30,410 Microsoft's own guidance is blunt about it, warning makers 241 00:07:30,410 --> 00:07:32,970 not to create a separate agent for every subtask. 242 00:07:32,970 --> 00:07:35,089 The reasons to split are narrower than they look. 243 00:07:35,089 --> 00:07:37,689 A subtask earns its own agent in three cases. 244 00:07:37,689 --> 00:07:40,089 It's complex enough to need its own tools or knowledge. 245 00:07:40,089 --> 00:07:43,089 It needs different access rules than the main agent, or several 246 00:07:43,089 --> 00:07:45,730 other agents will reuse it outside those cases. 247 00:07:45,730 --> 00:07:48,939 A child agent does the job with far less overhead. 248 00:07:48,939 --> 00:07:50,939 How a request moves through the system. 249 00:07:50,939 --> 00:07:53,300 Knowing the two shapes isn't the same as knowing what happens 250 00:07:53,300 --> 00:07:54,939 when a person types a question. 251 00:07:54,939 --> 00:07:58,339 So let's follow one request from start to finish. 252 00:07:58,339 --> 00:08:00,300 Someone asks something in plain language. 253 00:08:00,300 --> 00:08:01,500 Maybe in teams. 254 00:08:01,500 --> 00:08:03,699 The front door agent reads the request and decides which 255 00:08:03,699 --> 00:08:05,180 specialist fits it. 256 00:08:05,180 --> 00:08:07,540 Hands that specialist a task, waits for the answer and then 257 00:08:07,540 --> 00:08:09,500 combines whatever comes back into one response. 258 00:08:09,500 --> 00:08:11,819 Then it delivers that single response to the person. 259 00:08:11,819 --> 00:08:13,220 Microsoft is specific about that 260 00:08:13,220 --> 00:08:14,420 pattern, and the reason is worth 261 00:08:14,420 --> 00:08:15,180 understanding. 262 00:08:15,180 --> 00:08:16,420 The guidance tells the parent 263 00:08:16,420 --> 00:08:17,939 agent to invoke the specialists, 264 00:08:17,939 --> 00:08:19,180 wait for them, combine the 265 00:08:19,180 --> 00:08:20,740 findings, and send exactly one 266 00:08:20,740 --> 00:08:21,300 response. 267 00:08:21,300 --> 00:08:23,379 The company's own phrasing is memorable here. 268 00:08:23,379 --> 00:08:25,660 Subagents are researchers, not responders. 269 00:08:25,660 --> 00:08:28,819 There's a second benefit to that rule, and it's a quiet one when 270 00:08:28,819 --> 00:08:31,800 only the manager speaks, every answer passes through one voice, 271 00:08:31,800 --> 00:08:34,899 the tone stays steady, and the user never has to work out which 272 00:08:34,899 --> 00:08:36,059 agent said what. 273 00:08:36,059 --> 00:08:38,379 It also keeps the exchange easier to review later, because 274 00:08:38,379 --> 00:08:41,220 there's a single record of what the person actually saw. 275 00:08:41,220 --> 00:08:44,309 Skip that instruction and you get a specific annoying failure. 276 00:08:44,309 --> 00:08:45,529 The specialists start replying 277 00:08:45,529 --> 00:08:46,830 to the person directly, so the 278 00:08:46,830 --> 00:08:48,470 user sees two or three partial 279 00:08:48,470 --> 00:08:49,789 answers instead of one clean 280 00:08:49,789 --> 00:08:50,269 one. 281 00:08:50,269 --> 00:08:52,950 The fix is to tell every subagent in its own instructions 282 00:08:52,950 --> 00:08:54,149 that it's a subagent. 283 00:08:54,149 --> 00:08:56,909 Ask it to return findings instead of talking to the user. 284 00:08:56,909 --> 00:08:59,909 Descriptions carry more weight than beginners expect. 285 00:08:59,909 --> 00:09:01,389 The front door agent chooses a 286 00:09:01,389 --> 00:09:02,909 specialist by reading its name 287 00:09:02,909 --> 00:09:03,870 and description. 288 00:09:03,870 --> 00:09:04,950 So vague or identical 289 00:09:04,950 --> 00:09:06,149 descriptions cause wrong 290 00:09:06,149 --> 00:09:06,710 routing. 291 00:09:06,710 --> 00:09:08,269 If two specialists both say they 292 00:09:08,269 --> 00:09:09,769 help with questions, the manager 293 00:09:09,769 --> 00:09:10,750 is guessing right each 294 00:09:10,750 --> 00:09:11,750 description so it names the 295 00:09:11,750 --> 00:09:12,909 exact area that specialist 296 00:09:12,909 --> 00:09:13,470 covers. 297 00:09:13,470 --> 00:09:15,309 And one more detail catches people out. 298 00:09:15,309 --> 00:09:16,509 The child agent already has the 299 00:09:16,509 --> 00:09:17,830 whole conversation, so it never 300 00:09:17,830 --> 00:09:18,750 makes the user repeat 301 00:09:18,750 --> 00:09:19,590 themselves. 302 00:09:19,590 --> 00:09:21,629 The connected agent only sees that history when you've 303 00:09:21,629 --> 00:09:23,309 switched on the context setting. 304 00:09:23,309 --> 00:09:26,250 Leave it off by accident, and your specialist asks the user to 305 00:09:26,250 --> 00:09:28,149 repeat something they just said. 306 00:09:28,149 --> 00:09:29,710 Notice what this whole picture is. 307 00:09:29,710 --> 00:09:32,470 It's a routing problem dressed up in AI clothing. 308 00:09:32,470 --> 00:09:33,669 The intelligence is real, but 309 00:09:33,669 --> 00:09:35,070 the behavior you'll debug most 310 00:09:35,070 --> 00:09:36,230 often comes down to clear 311 00:09:36,230 --> 00:09:37,350 instructions and distinct 312 00:09:37,350 --> 00:09:37,909 descriptions. 313 00:09:37,909 --> 00:09:39,230 Get those right and the system 314 00:09:39,230 --> 00:09:40,669 feels calm and predictable, 315 00:09:40,669 --> 00:09:41,830 while vague ones make it feel 316 00:09:41,830 --> 00:09:43,250 like a team with no manager at 317 00:09:43,250 --> 00:09:43,850 all. 318 00:09:43,850 --> 00:09:44,970 Where this fits in Microsoft 319 00:09:44,970 --> 00:09:45,929 three hundred sixty five and 320 00:09:45,929 --> 00:09:46,450 Azure. 321 00:09:46,450 --> 00:09:48,730 Now for the two questions that decide whether any of this is 322 00:09:48,730 --> 00:09:50,769 real for you, where it lives, and what it costs. 323 00:09:50,769 --> 00:09:53,929 Copilot studio is its own web studio, but what you build shows 324 00:09:53,929 --> 00:09:56,169 up inside Microsoft three hundred sixty five. 325 00:09:56,169 --> 00:09:57,690 Agents can appear in teams in 326 00:09:57,690 --> 00:09:58,889 Microsoft Copilot and in 327 00:09:58,889 --> 00:10:00,490 SharePoint, and they can also be 328 00:10:00,490 --> 00:10:01,809 published to external channels 329 00:10:01,809 --> 00:10:03,169 like websites, apps, and social 330 00:10:03,169 --> 00:10:04,009 platforms. 331 00:10:04,009 --> 00:10:06,129 So the studio is separate, but the agents live where your 332 00:10:06,129 --> 00:10:06,889 people already are. 333 00:10:06,889 --> 00:10:09,129 You have two ways to pay, and the difference matters. 334 00:10:09,129 --> 00:10:10,370 A Microsoft three hundred sixty 335 00:10:10,370 --> 00:10:11,769 five copilot license costs 336 00:10:11,769 --> 00:10:13,409 thirty dollars per user per 337 00:10:13,409 --> 00:10:15,169 month, and it includes Copilot 338 00:10:15,169 --> 00:10:16,529 chat, plus the standard engine 339 00:10:16,529 --> 00:10:18,289 in Copilot Studio for building 340 00:10:18,289 --> 00:10:19,450 internal agents. 341 00:10:19,450 --> 00:10:22,490 If you need to publish agents to external channels or let people 342 00:10:22,490 --> 00:10:24,250 without that license use them. 343 00:10:24,250 --> 00:10:27,289 You step up to the standalone Copilot Studio license. 344 00:10:27,289 --> 00:10:29,250 Underneath, the license usage is metered. 345 00:10:29,250 --> 00:10:32,529 Microsoft sells capacity packs of twenty five thousand copilot 346 00:10:32,529 --> 00:10:34,809 credits for two hundred dollars per pack per month. 347 00:10:34,809 --> 00:10:36,289 There's also a pay as you go 348 00:10:36,289 --> 00:10:37,889 option that builds through an 349 00:10:37,889 --> 00:10:39,049 Azure subscription. 350 00:10:39,049 --> 00:10:41,700 You need that Azure subscription to run agents, and Microsoft 351 00:10:41,700 --> 00:10:43,779 confirms the capability is the same either way. 352 00:10:43,779 --> 00:10:45,100 Only the way you pay changes. 353 00:10:45,100 --> 00:10:46,659 Multi agent setups can cost more 354 00:10:46,659 --> 00:10:48,100 than people budget for, and the 355 00:10:48,100 --> 00:10:49,259 reason is how credits get 356 00:10:49,259 --> 00:10:49,779 charged. 357 00:10:49,779 --> 00:10:52,100 Credits are counted per action, not per conversation. 358 00:10:52,100 --> 00:10:53,980 A classic answer costs one credit. 359 00:10:53,980 --> 00:10:57,460 A generative answer costs two and an agent action costs five. 360 00:10:57,460 --> 00:10:58,779 Tenant graph grounding, which 361 00:10:58,779 --> 00:10:59,899 pulls in your organization's 362 00:10:59,899 --> 00:11:01,659 data for better answers, costs 363 00:11:01,659 --> 00:11:02,299 ten. 364 00:11:02,299 --> 00:11:05,019 So every extra handoff between agents adds reasoning and 365 00:11:05,019 --> 00:11:07,419 actions, and every one of those shows up on the bill. 366 00:11:07,419 --> 00:11:08,460 There's a friendly wrinkle here 367 00:11:08,460 --> 00:11:09,980 for internal use when someone 368 00:11:09,980 --> 00:11:10,259 with that. 369 00:11:10,259 --> 00:11:10,860 Microsoft three. 370 00:11:10,860 --> 00:11:13,580 Sixty five copilot license uses an agent inside Microsoft three 371 00:11:13,580 --> 00:11:14,580 hundred sixty five. 372 00:11:14,580 --> 00:11:17,659 A lot of that everyday usage is included at no extra charge. 373 00:11:17,659 --> 00:11:19,419 Fair use limits still apply. 374 00:11:19,419 --> 00:11:22,179 So internal helpers often cost far less than the sticker math 375 00:11:22,179 --> 00:11:25,019 suggests, and customer facing agents are where the meter 376 00:11:25,019 --> 00:11:26,419 really starts to move. 377 00:11:26,419 --> 00:11:28,580 Two more things are worth knowing before you commit. 378 00:11:28,580 --> 00:11:29,379 Unused credits. 379 00:11:29,379 --> 00:11:30,500 Don't carry over to the next 380 00:11:30,500 --> 00:11:32,019 month, so you can't bank them, 381 00:11:32,019 --> 00:11:33,500 and Microsoft enforces capacity 382 00:11:33,500 --> 00:11:34,659 at one hundred and twenty five 383 00:11:34,659 --> 00:11:35,700 percent, at which point custom 384 00:11:35,700 --> 00:11:37,000 agents get disabled and users 385 00:11:37,000 --> 00:11:38,220 see a message saying the agent 386 00:11:38,220 --> 00:11:39,470 is unavailable. 387 00:11:39,470 --> 00:11:42,190 There is a usage estimator to forecast this, and you can set a 388 00:11:42,190 --> 00:11:45,110 monthly limit per agent, so the surprise is avoidable if you 389 00:11:45,110 --> 00:11:46,029 plan for it. 390 00:11:46,029 --> 00:11:48,509 Azure shows up in one more place beyond the meter. 391 00:11:48,509 --> 00:11:49,629 When you want to grant answers 392 00:11:49,629 --> 00:11:50,710 in data that lives outside 393 00:11:50,710 --> 00:11:51,850 Microsoft three hundred sixty 394 00:11:51,850 --> 00:11:52,190 five. 395 00:11:52,190 --> 00:11:53,950 You often reach for services in Azure. 396 00:11:53,950 --> 00:11:56,590 That's why the two worlds usually get mentioned together. 397 00:11:56,590 --> 00:11:59,629 For a beginner, the practical takeaway is simpler than that. 398 00:11:59,629 --> 00:12:00,669 Microsoft three hundred sixty 399 00:12:00,669 --> 00:12:01,809 five is where the agent meets 400 00:12:01,809 --> 00:12:03,590 people, and Azure is where the 401 00:12:03,590 --> 00:12:04,830 usage gets paid for and the 402 00:12:04,830 --> 00:12:06,029 outside connections get plugged 403 00:12:06,029 --> 00:12:07,029 in. 404 00:12:07,029 --> 00:12:07,909 The mistakes that break 405 00:12:07,909 --> 00:12:09,629 multi-agent setups you've seen 406 00:12:09,629 --> 00:12:10,950 the mechanics, the platform, and 407 00:12:10,950 --> 00:12:11,389 the cost. 408 00:12:11,389 --> 00:12:12,429 Now let's name the mistakes, 409 00:12:12,429 --> 00:12:13,389 because these are the ones that 410 00:12:13,389 --> 00:12:15,070 turn a clean demo into a quiet 411 00:12:15,070 --> 00:12:15,710 failure. 412 00:12:15,710 --> 00:12:17,629 The first mistake is splitting too early. 413 00:12:17,629 --> 00:12:20,909 A two agent demo looks flawless, so teams add more agents and the 414 00:12:20,909 --> 00:12:22,350 setup starts to wobble. 415 00:12:22,350 --> 00:12:25,129 Microsoft's own advice is to start with one agent and to 416 00:12:25,129 --> 00:12:25,990 split only once. 417 00:12:25,990 --> 00:12:27,429 You can point to a real boundary. 418 00:12:27,429 --> 00:12:28,830 Fail to name that boundary and 419 00:12:28,830 --> 00:12:30,269 you're adding overhead for 420 00:12:30,269 --> 00:12:30,909 nothing. 421 00:12:30,909 --> 00:12:33,629 The second mistake is giving two specialists the same knowledge. 422 00:12:33,629 --> 00:12:36,590 If two subagents both search the same document library, one of 423 00:12:36,590 --> 00:12:37,879 them finds the answer first. 424 00:12:37,879 --> 00:12:40,799 The second one either duplicates it or skips its search entirely. 425 00:12:40,799 --> 00:12:43,960 Either way, you've paid for a specialist that adds nothing. 426 00:12:43,960 --> 00:12:45,120 Give each one a distinct source. 427 00:12:45,120 --> 00:12:47,399 And if you only have one source, don't split at all. 428 00:12:47,399 --> 00:12:50,399 The third mistake is leaving Subagents unaware of their role. 429 00:12:50,399 --> 00:12:52,519 If you don't tell a subagent, it's a subagent. 430 00:12:52,519 --> 00:12:53,759 It behaves like a standalone 431 00:12:53,759 --> 00:12:55,279 agent and replies to the user 432 00:12:55,279 --> 00:12:56,120 directly. 433 00:12:56,120 --> 00:12:58,879 The person then gets two answers to one question. 434 00:12:58,879 --> 00:13:00,519 Microsoft's guidance is to state 435 00:13:00,519 --> 00:13:01,840 the role in every subagents 436 00:13:01,840 --> 00:13:02,879 instructions. 437 00:13:02,879 --> 00:13:04,600 Use firm language like must and 438 00:13:04,600 --> 00:13:06,240 never because soft phrasing 439 00:13:06,240 --> 00:13:07,279 loses when instructions 440 00:13:07,279 --> 00:13:08,799 conflict. 441 00:13:08,799 --> 00:13:10,759 The fourth mistake is vague descriptions. 442 00:13:10,759 --> 00:13:12,879 The front door agent routes by description, so two agents 443 00:13:12,879 --> 00:13:15,320 described as helpful with questions will get misrouted. 444 00:13:15,320 --> 00:13:18,620 Name the exact area in every description, like searching HR 445 00:13:18,620 --> 00:13:21,679 policy documents or searching the IT knowledgebase. 446 00:13:21,679 --> 00:13:24,159 The fifth mistake is testing only the easy path. 447 00:13:24,159 --> 00:13:26,519 Microsoft recommends asking questions that match no 448 00:13:26,519 --> 00:13:29,840 specialist at all to see whether the system fails gracefully or 449 00:13:29,840 --> 00:13:31,039 invents an answer. 450 00:13:31,039 --> 00:13:32,279 Testing only the questions you 451 00:13:32,279 --> 00:13:33,440 know will work teaches you 452 00:13:33,440 --> 00:13:34,470 nothing about the day it 453 00:13:34,470 --> 00:13:35,299 matters. 454 00:13:35,299 --> 00:13:36,460 Then there's the honest part, 455 00:13:36,460 --> 00:13:37,700 and I'll label it as opinion, 456 00:13:37,700 --> 00:13:38,620 because it comes from one 457 00:13:38,620 --> 00:13:39,860 consultancy that builds these 458 00:13:39,860 --> 00:13:40,700 for clients. 459 00:13:40,700 --> 00:13:41,740 Their review argues that 460 00:13:41,740 --> 00:13:43,179 multi-agent orchestration works 461 00:13:43,179 --> 00:13:44,620 in a controlled demo with two 462 00:13:44,620 --> 00:13:46,340 agents in a real deployment, 463 00:13:46,340 --> 00:13:47,820 with five or six agents across 464 00:13:47,820 --> 00:13:48,820 business domains. 465 00:13:48,820 --> 00:13:50,139 They say it falls apart. 466 00:13:50,139 --> 00:13:52,460 There are specific complaints are brittle handoffs, context 467 00:13:52,460 --> 00:13:55,259 that doesn't transfer cleanly, and the manager agent that picks 468 00:13:55,259 --> 00:13:56,019 the wrong specialist. 469 00:13:56,019 --> 00:13:58,179 They add that fixing the routing means building your own 470 00:13:58,179 --> 00:13:59,539 orchestration on top. 471 00:13:59,539 --> 00:14:01,820 Way that carefully before you take it as settled. 472 00:14:01,820 --> 00:14:05,059 It's one practitioner's view and they sell a competing approach. 473 00:14:05,059 --> 00:14:06,419 So it's not the last word. 474 00:14:06,419 --> 00:14:07,340 But Microsoft's own 475 00:14:07,340 --> 00:14:08,899 documentation backs up two parts 476 00:14:08,899 --> 00:14:09,379 of it. 477 00:14:09,379 --> 00:14:11,820 Since Microsoft itself reports that connected agents add 478 00:14:11,820 --> 00:14:15,139 latency and create more surface to test, manage, and govern. 479 00:14:15,139 --> 00:14:17,700 The same review also warns that costs are hard to predict before 480 00:14:17,700 --> 00:14:20,539 real traffic arrives, with one estimate ballooning well past 481 00:14:20,539 --> 00:14:21,539 the initial projection. 482 00:14:21,539 --> 00:14:22,820 There's an affirmative side here 483 00:14:22,820 --> 00:14:23,980 because these are fixable 484 00:14:23,980 --> 00:14:25,399 problems rather than reasons to 485 00:14:25,399 --> 00:14:26,059 walk away. 486 00:14:26,059 --> 00:14:27,440 Clear instructions, distinct 487 00:14:27,440 --> 00:14:28,659 knowledge sources, accurate 488 00:14:28,659 --> 00:14:30,500 descriptions and honest testing 489 00:14:30,500 --> 00:14:31,700 solve most of the failures 490 00:14:31,700 --> 00:14:32,299 above. 491 00:14:32,299 --> 00:14:33,629 Teams that treat the manager 492 00:14:33,629 --> 00:14:35,350 agent as a routing problem and 493 00:14:35,350 --> 00:14:36,549 write its instructions like a 494 00:14:36,549 --> 00:14:38,190 job description, get predictable 495 00:14:38,190 --> 00:14:38,950 results. 496 00:14:38,950 --> 00:14:41,110 Teams that treat it as a magic box do not. 497 00:14:41,110 --> 00:14:43,870 None of this needs new tools or a bigger budget, and most of it 498 00:14:43,870 --> 00:14:47,029 comes down to writing things down clearly before you build. 499 00:14:47,029 --> 00:14:48,870 A quick recap and a question for you. 500 00:14:48,870 --> 00:14:49,990 So let's pull it together. 501 00:14:49,990 --> 00:14:51,710 Multi-Agent isn't a swarm. 502 00:14:51,710 --> 00:14:54,269 It's one agent playing manager with a few others doing narrow 503 00:14:54,269 --> 00:14:57,389 jobs, and child agents keep that split simple while connected 504 00:14:57,389 --> 00:15:00,389 agents add reach along with extra latency and cost. 505 00:15:00,389 --> 00:15:03,110 The rule to carry with you is this start with a single agent 506 00:15:03,110 --> 00:15:04,149 and split only once. 507 00:15:04,149 --> 00:15:05,429 You can name the boundary. 508 00:15:05,429 --> 00:15:08,149 That one rule will save you more money and more frustration than 509 00:15:08,149 --> 00:15:10,870 any clever architecture will, because it keeps you honest 510 00:15:10,870 --> 00:15:13,110 about why each agent exists. 511 00:15:13,110 --> 00:15:15,710 Before you go, I want to leave you with one question. 512 00:15:15,710 --> 00:15:17,429 If you had to hand one job to an 513 00:15:17,429 --> 00:15:18,950 agent tomorrow, what would it 514 00:15:18,950 --> 00:15:19,549 be? 515 00:15:19,549 --> 00:15:20,750 Think about the task that eats 516 00:15:20,750 --> 00:15:21,870 your week and follows the same 517 00:15:21,870 --> 00:15:23,149 steps every time, because that's 518 00:15:23,149 --> 00:15:24,309 almost always the right first 519 00:15:24,309 --> 00:15:24,750 answer. 520 00:15:24,750 --> 00:15:26,110 If this made the topic feel less 521 00:15:26,110 --> 00:15:28,450 intimidating, subscribe to M365 522 00:15:28,450 --> 00:15:30,070 FM to master the tools and the 523 00:15:30,070 --> 00:15:31,590 mindset shaping the future of 524 00:15:31,590 --> 00:15:32,230 work.