1
00:00:00,040 --> 00:00:01,600
Microsoft's Foundry catalog now
2
00:00:01,600 --> 00:00:03,020
holds more than ten thousand AI
3
00:00:03,020 --> 00:00:03,720
models.
4
00:00:03,720 --> 00:00:06,080
That number only matters if you
can turn a model into something
5
00:00:06,080 --> 00:00:07,160
that actually does work.
6
00:00:07,160 --> 00:00:08,160
You might assume that means
7
00:00:08,160 --> 00:00:09,800
hiring developers and setting up
8
00:00:09,800 --> 00:00:10,439
servers.
9
00:00:10,439 --> 00:00:12,060
Foundry agent service is how
10
00:00:12,060 --> 00:00:13,599
Microsoft lets you build those
11
00:00:13,599 --> 00:00:15,240
working agents without the
12
00:00:15,240 --> 00:00:16,719
server headache.
13
00:00:16,719 --> 00:00:20,120
Welcome to the M365 FM podcast,
where we explore the people,
14
00:00:20,120 --> 00:00:22,600
ideas, and technologies shaping
the future of work.
15
00:00:22,600 --> 00:00:24,679
By the end, you'll know what it
is, how it works, and whether
16
00:00:24,679 --> 00:00:26,039
you could use it.
17
00:00:26,039 --> 00:00:27,960
What an AI agent actually is.
18
00:00:27,960 --> 00:00:30,440
Start with the thing you already
know because it makes the rest
19
00:00:30,440 --> 00:00:31,440
of this easier.
20
00:00:31,440 --> 00:00:33,920
A chatbot waits for your
question, gives you an answer,
21
00:00:33,920 --> 00:00:35,439
and then stops right there.
22
00:00:35,439 --> 00:00:36,719
An agent takes your request and
23
00:00:36,719 --> 00:00:37,840
keeps working so it can look
24
00:00:37,840 --> 00:00:38,439
something up.
25
00:00:38,439 --> 00:00:40,719
Change a record, and then tell
you what it did.
26
00:00:40,719 --> 00:00:43,439
Picture a helper who handles
your travel booking for you.
27
00:00:43,439 --> 00:00:45,840
You say you need a trip next
month and the helper checks your
28
00:00:45,840 --> 00:00:49,119
calendar, finds a flight, books
it, and sends you the details.
29
00:00:49,119 --> 00:00:51,359
A chatbot would only describe
how to book that flight while an
30
00:00:51,359 --> 00:00:52,719
agent goes ahead and books it.
31
00:00:52,719 --> 00:00:55,479
Plane automation is different
again, and it's worth keeping
32
00:00:55,479 --> 00:00:56,439
the two apart.
33
00:00:56,439 --> 00:00:57,920
Automation follows a fixed set
34
00:00:57,920 --> 00:00:59,439
of steps every single time, so
35
00:00:59,439 --> 00:01:00,560
it breaks the moment something
36
00:01:00,560 --> 00:01:01,979
unexpected happens.
37
00:01:01,979 --> 00:01:05,060
An agent decides which step to
take next based on whatever it
38
00:01:05,060 --> 00:01:07,659
finds along the way, which is
why it holds up better when the
39
00:01:07,659 --> 00:01:08,859
real world gets messy.
40
00:01:08,859 --> 00:01:11,420
If that sounds like a small
difference, think about how much
41
00:01:11,420 --> 00:01:13,739
of your day disappears into
chasing things down.
42
00:01:13,739 --> 00:01:14,980
Someone asks for an update and
43
00:01:14,980 --> 00:01:16,719
you open three systems to
44
00:01:16,719 --> 00:01:17,340
answer.
45
00:01:17,340 --> 00:01:18,859
An agent can do that opening and
46
00:01:18,859 --> 00:01:20,299
checking for you, and then hand
47
00:01:20,299 --> 00:01:21,260
you the answer with the work
48
00:01:21,260 --> 00:01:22,099
already done.
49
00:01:22,099 --> 00:01:24,019
Every agent needs three simple
parts.
50
00:01:24,019 --> 00:01:25,420
The first part is a model, which
51
00:01:25,420 --> 00:01:26,540
is the AI that reads your
52
00:01:26,540 --> 00:01:27,780
request and decides what to do
53
00:01:27,780 --> 00:01:28,459
next.
54
00:01:28,459 --> 00:01:29,900
The second part is instructions,
55
00:01:29,900 --> 00:01:31,140
which act like a job description
56
00:01:31,140 --> 00:01:32,299
and tell the agent what it is
57
00:01:32,299 --> 00:01:32,900
for.
58
00:01:32,900 --> 00:01:35,900
The third part is tools, which
are the ways the agent can act,
59
00:01:35,900 --> 00:01:38,099
such as searching the web or
updating a record.
60
00:01:38,099 --> 00:01:39,420
Put those three parts together
61
00:01:39,420 --> 00:01:40,659
and you get something that can
62
00:01:40,659 --> 00:01:42,019
work through a task in several
63
00:01:42,019 --> 00:01:43,299
steps instead of stopping at one
64
00:01:43,299 --> 00:01:44,019
reply.
65
00:01:44,019 --> 00:01:45,260
That's the whole idea behind
66
00:01:45,260 --> 00:01:46,700
every agent you'll meet today,
67
00:01:46,700 --> 00:01:47,780
no matter how big the name
68
00:01:47,780 --> 00:01:48,500
sounds.
69
00:01:48,500 --> 00:01:50,180
The model, it turns out, is the
easy part.
70
00:01:50,180 --> 00:01:51,379
Everything around it is where
71
00:01:51,379 --> 00:01:52,939
the real work lives, and that
72
00:01:52,939 --> 00:01:54,099
idea comes back in a few
73
00:01:54,099 --> 00:01:55,180
minutes.
74
00:01:55,180 --> 00:01:57,299
What Microsoft Founder Agent
service is.
75
00:01:57,299 --> 00:01:58,900
Now let's put a name on the
product.
76
00:01:58,900 --> 00:02:01,704
Microsoft Foundry is the
platform, and Foundry agent
77
00:02:01,704 --> 00:02:04,430
service is the part of it built
for creating and running agents.
78
00:02:04,430 --> 00:02:05,670
It lives in Azure, which is
79
00:02:05,670 --> 00:02:07,150
simply Microsoft's cloud, the
80
00:02:07,150 --> 00:02:08,189
place where companies rent
81
00:02:08,189 --> 00:02:09,310
computing power instead of
82
00:02:09,310 --> 00:02:10,469
buying and maintaining their own
83
00:02:10,469 --> 00:02:11,270
machines.
84
00:02:11,270 --> 00:02:13,189
The name is a small clue as
well.
85
00:02:13,189 --> 00:02:16,189
A foundry is a place where raw
material goes in and finished.
86
00:02:16,189 --> 00:02:18,710
Things come out, and that's
roughly the idea here.
87
00:02:18,710 --> 00:02:19,990
Models and tools go in and
88
00:02:19,990 --> 00:02:21,069
working agents come out the
89
00:02:21,069 --> 00:02:21,830
other side.
90
00:02:21,830 --> 00:02:23,669
The word that matters most here
is managed.
91
00:02:23,669 --> 00:02:24,830
When something is managed,
92
00:02:24,830 --> 00:02:26,189
Microsoft looks after the
93
00:02:26,189 --> 00:02:27,409
servers, the scaling and the
94
00:02:27,409 --> 00:02:28,110
updates.
95
00:02:28,110 --> 00:02:29,389
So you spend your time on the
96
00:02:29,389 --> 00:02:30,430
agent instead of on the
97
00:02:30,430 --> 00:02:30,990
plumbing.
98
00:02:30,990 --> 00:02:32,189
You still pay for what you use,
99
00:02:32,189 --> 00:02:33,370
but you don't sit up at two in
100
00:02:33,370 --> 00:02:34,349
the morning restarting a
101
00:02:34,349 --> 00:02:34,830
machine.
102
00:02:34,830 --> 00:02:37,349
That single word explains why a
beginner can get something.
103
00:02:37,349 --> 00:02:40,310
Working in an afternoon foundry
also meets you at whatever level
104
00:02:40,310 --> 00:02:41,389
you're comfortable with.
105
00:02:41,389 --> 00:02:44,750
At one end, you build an agent
with nothing but settings and
106
00:02:44,750 --> 00:02:47,909
plain written instructions,
which means no coding at all.
107
00:02:47,909 --> 00:02:50,710
At the other end, you bring your
own code and your own framework
108
00:02:50,710 --> 00:02:52,310
and foundry runs it for you.
109
00:02:52,310 --> 00:02:53,990
You can start at the simple end
110
00:02:53,990 --> 00:02:55,310
and move toward the code end
111
00:02:55,310 --> 00:02:56,909
later without throwing away what
112
00:02:56,909 --> 00:02:57,629
you built.
113
00:02:57,629 --> 00:02:59,870
There are three kinds of agent
to know about, and the names are
114
00:02:59,870 --> 00:03:01,189
friendlier than they look.
115
00:03:01,189 --> 00:03:04,169
A prompt agent is built from
settings and instructions, and
116
00:03:04,169 --> 00:03:06,090
it's the fastest place to begin.
117
00:03:06,090 --> 00:03:07,210
A voice agent is built for
118
00:03:07,210 --> 00:03:08,810
spoken conversations, so it can
119
00:03:08,810 --> 00:03:09,969
answer on a phone line or in a
120
00:03:09,969 --> 00:03:10,689
browser.
121
00:03:10,689 --> 00:03:13,770
A hosted agent is your own code,
packaged up and run by foundry,
122
00:03:13,770 --> 00:03:16,009
with the servers and the scaling
handled for you.
123
00:03:16,009 --> 00:03:18,330
Most people start with the first
kind and never need the other
124
00:03:18,330 --> 00:03:19,889
two, which is perfectly fine.
125
00:03:19,889 --> 00:03:21,289
Underneath all three, foundry
126
00:03:21,289 --> 00:03:23,169
brings agents, models, and tools
127
00:03:23,169 --> 00:03:24,650
together in one place with
128
00:03:24,650 --> 00:03:25,969
shared access control and shared
129
00:03:25,969 --> 00:03:26,849
monitoring.
130
00:03:26,849 --> 00:03:28,289
In practice, that means one
131
00:03:28,289 --> 00:03:29,629
place to sign in, one place to
132
00:03:29,629 --> 00:03:31,090
set permissions, and one place
133
00:03:31,090 --> 00:03:32,250
to watch what your agents are
134
00:03:32,250 --> 00:03:32,689
doing.
135
00:03:32,689 --> 00:03:33,210
That matters.
136
00:03:33,210 --> 00:03:36,169
Later, when you want to know who
can touch an agent and what that
137
00:03:36,169 --> 00:03:37,650
agent has been doing.
138
00:03:37,650 --> 00:03:38,729
Models the brain.
139
00:03:38,729 --> 00:03:39,610
You can swap.
140
00:03:39,610 --> 00:03:41,169
Every agent runs on a model, and
141
00:03:41,169 --> 00:03:42,610
the model is the part that does
142
00:03:42,610 --> 00:03:43,330
the thinking.
143
00:03:43,330 --> 00:03:44,689
Think of it as the brain of the
144
00:03:44,689 --> 00:03:46,550
agent because it decides what to
145
00:03:46,550 --> 00:03:48,650
say or do while the tools handle
146
00:03:48,650 --> 00:03:49,370
the acting.
147
00:03:49,370 --> 00:03:51,610
The scale here is bigger than
most people expect.
148
00:03:51,610 --> 00:03:54,610
The foundry catalog offers more
than ten thousand models built
149
00:03:54,610 --> 00:03:58,090
by Microsoft, OpenAI, anthropic,
meta, and plenty of others.
150
00:03:58,090 --> 00:03:59,569
Having that many options sounds
151
00:03:59,569 --> 00:04:01,129
overwhelming, but you don't have
152
00:04:01,129 --> 00:04:02,280
to shop around for long, because
153
00:04:02,280 --> 00:04:03,509
you can start with a sensible
154
00:04:03,509 --> 00:04:04,789
default and change your mind
155
00:04:04,789 --> 00:04:05,349
later.
156
00:04:05,349 --> 00:04:08,030
The useful part for you is that
you can swap the model without
157
00:04:08,030 --> 00:04:09,189
rewriting your agent.
158
00:04:09,189 --> 00:04:12,069
If a cheap, fast model does the
job, you use that one.
159
00:04:12,069 --> 00:04:14,550
If a harder task needs a
stronger model, you change a
160
00:04:14,550 --> 00:04:16,949
single setting and keep
everything else exactly as it
161
00:04:16,949 --> 00:04:19,829
was that freedom matters because
it means you aren't locked into
162
00:04:19,829 --> 00:04:21,110
a choice you made on day one.
163
00:04:21,110 --> 00:04:22,350
Different jobs suit different
164
00:04:22,350 --> 00:04:23,350
models in the same way that
165
00:04:23,350 --> 00:04:24,910
different errands suit different
166
00:04:24,910 --> 00:04:25,790
vehicles.
167
00:04:25,790 --> 00:04:28,269
A quick reply is a small errand,
and a long chain of careful
168
00:04:28,269 --> 00:04:29,509
reasoning is a big one.
169
00:04:29,509 --> 00:04:30,709
The model you pick should match
170
00:04:30,709 --> 00:04:31,870
the weight of the task in front
171
00:04:31,870 --> 00:04:32,350
of it.
172
00:04:32,350 --> 00:04:35,149
The catalog also spans a wide
range, from tiny models that
173
00:04:35,149 --> 00:04:37,870
cost almost nothing to large
ones that reason through hard
174
00:04:37,870 --> 00:04:39,509
problems step by step.
175
00:04:39,509 --> 00:04:41,870
You don't need to memorize any
of it, because the point is
176
00:04:41,870 --> 00:04:44,509
simply that you have options and
you can change your mind without
177
00:04:44,509 --> 00:04:45,829
rebuilding the agent.
178
00:04:45,829 --> 00:04:48,189
So you have a brain you can
swap, and that's genuinely
179
00:04:48,189 --> 00:04:49,069
useful on its own.
180
00:04:49,069 --> 00:04:51,670
The trouble is that a brain
can't lift a finger by itself,
181
00:04:51,670 --> 00:04:54,550
and that gap is exactly what the
next part fills.
182
00:04:54,550 --> 00:04:55,670
Tools and knowledge.
183
00:04:55,670 --> 00:04:57,310
How agents take action.
184
00:04:57,310 --> 00:04:58,790
A model on its own can only
185
00:04:58,790 --> 00:05:00,430
talk, and talking isn't the same
186
00:05:00,430 --> 00:05:01,189
as doing.
187
00:05:01,189 --> 00:05:03,129
To do real work, an agent needs
188
00:05:03,129 --> 00:05:04,730
tools and a tool is simply a
189
00:05:04,730 --> 00:05:05,970
capability the agent is allowed
190
00:05:05,970 --> 00:05:06,370
to use.
191
00:05:06,370 --> 00:05:09,250
If the model is the brain, the
tools are the hands.
192
00:05:09,250 --> 00:05:12,009
Foundry ships with built in
tools that you can switch on.
193
00:05:12,009 --> 00:05:13,569
Web search looks things up.
194
00:05:13,569 --> 00:05:16,730
File search reads your documents
and a code interpreter runs
195
00:05:16,730 --> 00:05:18,370
calculations and small programs.
196
00:05:18,370 --> 00:05:21,449
Memory is the forth and it lets
the agent hold on to useful
197
00:05:21,449 --> 00:05:22,769
details between conversations.
198
00:05:22,769 --> 00:05:25,290
These all come ready to use so
you get real value before you
199
00:05:25,290 --> 00:05:26,529
connect anything of your own.
200
00:05:26,529 --> 00:05:27,610
You aren't limited to those
201
00:05:27,610 --> 00:05:28,689
because you can add your own
202
00:05:28,689 --> 00:05:29,610
tools too.
203
00:05:29,610 --> 00:05:30,889
You can connect a tool through a
204
00:05:30,889 --> 00:05:32,389
function, through an open API
205
00:05:32,389 --> 00:05:34,689
definition or through MCP, which
206
00:05:34,689 --> 00:05:36,290
is a shared plug standard that
207
00:05:36,290 --> 00:05:38,050
lets AI tools connect to other
208
00:05:38,050 --> 00:05:38,889
systems.
209
00:05:38,889 --> 00:05:41,810
The exact labels matter less
than the idea, which is that
210
00:05:41,810 --> 00:05:44,889
your agent can reach the systems
your business already runs on.
211
00:05:44,889 --> 00:05:46,529
Knowledge is the other half of
the story.
212
00:05:46,529 --> 00:05:47,649
You can point an agent at your
213
00:05:47,649 --> 00:05:49,290
own content, such as files in
214
00:05:49,290 --> 00:05:50,930
SharePoint, data in Microsoft
215
00:05:50,930 --> 00:05:52,689
Fabric, or current results from
216
00:05:52,689 --> 00:05:53,490
Bing Search.
217
00:05:53,490 --> 00:05:56,490
When an agent answers from your
real data instead of guessing.
218
00:05:56,490 --> 00:05:57,410
That's called grounding.
219
00:05:57,410 --> 00:05:58,490
And it's the difference between
220
00:05:58,490 --> 00:06:00,209
a confident answer and a correct
221
00:06:00,209 --> 00:06:00,730
one.
222
00:06:00,730 --> 00:06:03,500
One more piece is worth knowing,
and it's called a toolbox.
223
00:06:03,500 --> 00:06:04,779
A toolbox lets you gather your
224
00:06:04,779 --> 00:06:06,579
tools once, expose them through
225
00:06:06,579 --> 00:06:08,259
a single connection, and reuse
226
00:06:08,259 --> 00:06:09,939
them across many agents with
227
00:06:09,939 --> 00:06:10,980
shared sign in and shared
228
00:06:10,980 --> 00:06:11,779
controls.
229
00:06:11,779 --> 00:06:12,980
It saves you from wiring the
230
00:06:12,980 --> 00:06:14,420
same tool into every new agent
231
00:06:14,420 --> 00:06:15,100
you build.
232
00:06:15,100 --> 00:06:16,459
To make this concrete picture an
233
00:06:16,459 --> 00:06:17,620
agent that helps onboard a new
234
00:06:17,620 --> 00:06:18,420
employee.
235
00:06:18,420 --> 00:06:19,860
In one request, it could pull
236
00:06:19,860 --> 00:06:21,120
the onboarding guidance from a
237
00:06:21,120 --> 00:06:22,420
knowledge base, create the
238
00:06:22,420 --> 00:06:24,019
account, draft a welcome email,
239
00:06:24,019 --> 00:06:25,500
and post a message in the team's
240
00:06:25,500 --> 00:06:26,180
channel.
241
00:06:26,180 --> 00:06:28,980
That's several tools and several
systems working together behind
242
00:06:28,980 --> 00:06:30,019
a single request.
243
00:06:30,019 --> 00:06:32,620
And it's the moment an agent
stops being a novelty and starts
244
00:06:32,620 --> 00:06:34,060
being genuinely useful.
245
00:06:34,060 --> 00:06:36,220
Now you can see why the model
was the easy part.
246
00:06:36,220 --> 00:06:37,500
The work sits in the tools, the
247
00:06:37,500 --> 00:06:38,939
data, and the permissions around
248
00:06:38,939 --> 00:06:40,300
them, and a toolbox is
249
00:06:40,300 --> 00:06:41,699
Microsoft's way of making that
250
00:06:41,699 --> 00:06:41,980
work.
251
00:06:41,980 --> 00:06:46,019
Manageable threads, runs,
conversations and responses.
252
00:06:46,019 --> 00:06:48,980
Memory is what makes an agent
feel less like a search box and
253
00:06:48,980 --> 00:06:51,699
more like a colleague, and it
comes down to a few terms that
254
00:06:51,699 --> 00:06:53,139
used to confuse everyone.
255
00:06:53,139 --> 00:06:54,139
The good news is that those
256
00:06:54,139 --> 00:06:55,180
terms are getting simpler, so
257
00:06:55,180 --> 00:06:56,339
this part is easier than it
258
00:06:56,339 --> 00:06:57,180
sounds.
259
00:06:57,180 --> 00:06:58,660
In the older way of doing
things.
260
00:06:58,660 --> 00:07:00,420
A conversation was called a
thread.
261
00:07:00,420 --> 00:07:03,149
A thread held your messages,
which were the back and forth
262
00:07:03,149 --> 00:07:04,360
between you and the agent.
263
00:07:04,360 --> 00:07:07,079
Each time the agent took a turn,
that was called a run, and you
264
00:07:07,079 --> 00:07:09,720
could watch the run move from
waiting to working to finished.
265
00:07:09,720 --> 00:07:12,519
The new away keeps the same idea
and changes the names.
266
00:07:12,519 --> 00:07:15,240
A conversation now holds the
history and a response is what
267
00:07:15,240 --> 00:07:18,000
the agent produces, which can be
an answer or an action.
268
00:07:18,000 --> 00:07:21,800
So threads became conversations
and runs became responses.
269
00:07:21,800 --> 00:07:23,959
And that's the whole rename in a
single line.
270
00:07:23,959 --> 00:07:25,879
The change does more than tidy
up the wording.
271
00:07:25,879 --> 00:07:27,160
A conversation can store more
272
00:07:27,160 --> 00:07:28,600
than messages because it can
273
00:07:28,600 --> 00:07:30,120
also keep track of the tools the
274
00:07:30,120 --> 00:07:31,240
agent called and what those
275
00:07:31,240 --> 00:07:32,240
tools returned.
276
00:07:32,240 --> 00:07:34,680
The agent ends up with a richer
memory of what happened, which
277
00:07:34,680 --> 00:07:36,639
makes it better at picking up
where you left off.
278
00:07:36,639 --> 00:07:38,199
History matters for a practical
reason.
279
00:07:38,199 --> 00:07:39,800
Most useful tasks take more than
280
00:07:39,800 --> 00:07:41,439
one message, and an agent that
281
00:07:41,439 --> 00:07:42,639
forgets the first half of a
282
00:07:42,639 --> 00:07:43,899
request isn't much use to
283
00:07:43,899 --> 00:07:44,519
anyone.
284
00:07:44,519 --> 00:07:47,360
Keeping the conversation lets
the agent follow a thread of
285
00:07:47,360 --> 00:07:49,720
thought across many turns the
way a person would.
286
00:07:49,720 --> 00:07:52,759
You felt this in every good chat
tool you've used, you ask a
287
00:07:52,759 --> 00:07:55,360
follow up question without
repeating yourself, and the tool
288
00:07:55,360 --> 00:07:58,160
understands because it kept the
earlier turns in view.
289
00:07:58,160 --> 00:07:59,439
Conversations give an agent that
290
00:07:59,439 --> 00:08:00,740
same ability, which is why they
291
00:08:00,740 --> 00:08:01,800
sit at the center of how it
292
00:08:01,800 --> 00:08:02,639
works.
293
00:08:02,639 --> 00:08:03,899
One practical note, since you
294
00:08:03,899 --> 00:08:05,660
may have read older guides, the
295
00:08:05,660 --> 00:08:07,060
old threads and runs setup is
296
00:08:07,060 --> 00:08:08,300
being retired in favor of this
297
00:08:08,300 --> 00:08:08,980
newer one.
298
00:08:08,980 --> 00:08:10,160
So if you're learning today,
299
00:08:10,160 --> 00:08:11,420
learn the words, conversation
300
00:08:11,420 --> 00:08:12,339
and response.
301
00:08:12,339 --> 00:08:14,180
That's where the platform has
moved.
302
00:08:14,180 --> 00:08:15,379
Hosting and scaling.
303
00:08:15,379 --> 00:08:16,660
Where the agent lives.
304
00:08:16,660 --> 00:08:17,980
An agent has to run somewhere
305
00:08:17,980 --> 00:08:19,220
and that somewhere has to stay
306
00:08:19,220 --> 00:08:20,620
available when people actually
307
00:08:20,620 --> 00:08:20,980
use it.
308
00:08:20,980 --> 00:08:21,819
That's what hosting means.
309
00:08:21,819 --> 00:08:24,939
And it's one of those jobs you'd
much rather hand to someone else
310
00:08:24,939 --> 00:08:25,540
for a prompt.
311
00:08:25,540 --> 00:08:27,379
Agent hosting is barely a topic
at all.
312
00:08:27,379 --> 00:08:30,000
Microsoft runs it for you, so
there's nothing to install,
313
00:08:30,000 --> 00:08:31,060
patch, or restart.
314
00:08:31,060 --> 00:08:33,299
You write the instructions,
choose the tools, and the agent
315
00:08:33,299 --> 00:08:35,820
is simply there when someone
calls it hosted agents need a
316
00:08:35,820 --> 00:08:38,820
little more explanation because
you're bringing your own code.
317
00:08:38,820 --> 00:08:40,340
Foundry runs that code in its
318
00:08:40,340 --> 00:08:41,740
own isolated space for each
319
00:08:41,740 --> 00:08:43,460
session, which is a tidy way of
320
00:08:43,460 --> 00:08:44,899
saying that one person's work
321
00:08:44,899 --> 00:08:46,059
stays separate from everyone
322
00:08:46,059 --> 00:08:46,740
else's.
323
00:08:46,740 --> 00:08:48,059
When more people use the agent,
324
00:08:48,059 --> 00:08:49,539
it scales up on its own, and
325
00:08:49,539 --> 00:08:51,179
when fewer do, it scales back
326
00:08:51,179 --> 00:08:51,700
down.
327
00:08:51,700 --> 00:08:54,460
The idle behavior is worth
pausing on because it answers
328
00:08:54,460 --> 00:08:55,820
the cost question.
329
00:08:55,820 --> 00:08:58,779
When nobody is using an agent,
it goes to sleep and its state
330
00:08:58,779 --> 00:08:59,740
is kept safe.
331
00:08:59,740 --> 00:09:02,620
When the next person arrives, it
wakes up exactly where it left
332
00:09:02,620 --> 00:09:05,950
off, and you aren't paying for
the quiet hours in between.
333
00:09:05,950 --> 00:09:09,190
Every hosted agent also gets its
own identity and its own address
334
00:09:09,190 --> 00:09:11,789
automatically, and that sets up
the next part of the story.
335
00:09:11,789 --> 00:09:14,950
That identity is how the agent
proves who it is when it reaches
336
00:09:14,950 --> 00:09:15,909
for your systems.
337
00:09:15,909 --> 00:09:18,190
One more detail is worth
knowing, even if it stays
338
00:09:18,190 --> 00:09:19,590
invisible to you.
339
00:09:19,590 --> 00:09:22,350
Hosted agents run in a long list
of Azure regions around the
340
00:09:22,350 --> 00:09:24,909
world, so a company can often
keep its agents close to where
341
00:09:24,909 --> 00:09:26,509
its data already lives.
342
00:09:26,509 --> 00:09:28,750
For the audience here, the short
version is that scaling is a
343
00:09:28,750 --> 00:09:31,230
problem the platform has already
solved for you.
344
00:09:31,230 --> 00:09:33,269
Safety, identity and evaluation.
345
00:09:33,269 --> 00:09:34,389
Once an agent can reach into
346
00:09:34,389 --> 00:09:35,950
your systems, trust becomes the
347
00:09:35,950 --> 00:09:36,629
whole game.
348
00:09:36,629 --> 00:09:39,350
And this is where Managed
platform earns its place.
349
00:09:39,350 --> 00:09:40,549
Start with identity.
350
00:09:40,549 --> 00:09:42,309
Every agent can get its own sign
351
00:09:42,309 --> 00:09:43,669
in, which means it only touches
352
00:09:43,669 --> 00:09:44,909
the things you've allowed and
353
00:09:44,909 --> 00:09:45,750
nothing else.
354
00:09:45,750 --> 00:09:47,169
You can also see what it did so
355
00:09:47,169 --> 00:09:48,789
an agent stays accountable in a
356
00:09:48,789 --> 00:09:50,110
way a shared password never
357
00:09:50,110 --> 00:09:50,669
could.
358
00:09:50,669 --> 00:09:52,029
Guardrails are the second layer.
359
00:09:52,029 --> 00:09:54,909
These are built in filters that
help block unsafe output.
360
00:09:54,909 --> 00:09:56,110
They also reduce the risk of
361
00:09:56,110 --> 00:09:57,190
someone tricking the agent
362
00:09:57,190 --> 00:09:58,470
through a prompt, which is the
363
00:09:58,470 --> 00:09:59,909
practice of slipping hidden
364
00:09:59,909 --> 00:10:01,230
instructions into something the
365
00:10:01,230 --> 00:10:02,230
agent reads.
366
00:10:02,230 --> 00:10:04,889
You can also keep an agent
inside your own private network
367
00:10:04,889 --> 00:10:07,370
so its data doesn't wander
anywhere you haven't approved.
368
00:10:07,370 --> 00:10:09,610
Evaluation is the part that
surprises people because it
369
00:10:09,610 --> 00:10:11,009
works a lot like testing.
370
00:10:11,009 --> 00:10:12,610
Foundry includes checkers that
score.
371
00:10:12,610 --> 00:10:13,690
Whether the agent finished the
372
00:10:13,690 --> 00:10:15,129
task, whether it followed your
373
00:10:15,129 --> 00:10:16,610
instructions, and whether it
374
00:10:16,610 --> 00:10:18,009
used the right tools along the
375
00:10:18,009 --> 00:10:18,690
way.
376
00:10:18,690 --> 00:10:21,490
You can think of them as unit
tests for an agent, and they run
377
00:10:21,490 --> 00:10:23,370
before real people have to
depend on it.
378
00:10:23,370 --> 00:10:27,090
Observability is the last piece,
and it simply means you can
379
00:10:27,090 --> 00:10:28,850
trace every step the agent took.
380
00:10:28,850 --> 00:10:29,889
When something goes wrong, you
381
00:10:29,889 --> 00:10:30,909
follow the trail instead of
382
00:10:30,909 --> 00:10:31,250
guessing.
383
00:10:31,250 --> 00:10:33,129
And that's how confidence builds
over time.
384
00:10:33,129 --> 00:10:35,490
None of this means you can
switch off your own judgment.
385
00:10:35,490 --> 00:10:38,090
A managed platform gives you the
controls, but you still decide
386
00:10:38,090 --> 00:10:41,169
what an agent is allowed to do
and how far it's trusted.
387
00:10:41,169 --> 00:10:42,389
The sensible approach is to
388
00:10:42,389 --> 00:10:43,750
start with a small, low risk
389
00:10:43,750 --> 00:10:44,490
task.
390
00:10:44,490 --> 00:10:45,929
Watch how the agent behaves and
391
00:10:45,929 --> 00:10:47,490
widen its reach only as your
392
00:10:47,490 --> 00:10:48,409
confidence grows.
393
00:10:48,409 --> 00:10:50,970
When an agent is ready, you can
publish it to Microsoft Teams
394
00:10:50,970 --> 00:10:53,090
and Microsoft three hundred
sixty five copilot, so people
395
00:10:53,090 --> 00:10:55,289
meet it in the tools they
already use every day.
396
00:10:55,289 --> 00:10:57,730
That's what separates a clever
experiment from something your
397
00:10:57,730 --> 00:10:59,250
team actually relies on.
398
00:10:59,250 --> 00:11:01,210
How it relates to Copilot
Studio.
399
00:11:01,210 --> 00:11:02,450
By now, a fair question is
400
00:11:02,450 --> 00:11:03,889
forming because Microsoft has
401
00:11:03,889 --> 00:11:05,110
more than one way to build an
402
00:11:05,110 --> 00:11:06,529
agent and the names start to
403
00:11:06,529 --> 00:11:07,470
blur together.
404
00:11:07,470 --> 00:11:08,309
Let's clear that up.
405
00:11:08,309 --> 00:11:10,389
Copilot studio is the low code
door.
406
00:11:10,389 --> 00:11:11,629
It's built for business users
407
00:11:11,629 --> 00:11:12,909
and makers, and it connects
408
00:11:12,909 --> 00:11:14,149
quickly to the Microsoft three
409
00:11:14,149 --> 00:11:15,309
hundred sixty five tools and
410
00:11:15,309 --> 00:11:16,590
power platform connectors your
411
00:11:16,590 --> 00:11:18,269
company already owns for many
412
00:11:18,269 --> 00:11:18,789
teams.
413
00:11:18,789 --> 00:11:20,990
It's the fastest route to a
working agent.
414
00:11:20,990 --> 00:11:22,429
If your goal is a helpful
assistant.
415
00:11:22,429 --> 00:11:24,870
Inside teams, this is usually
where you start.
416
00:11:24,870 --> 00:11:26,509
Foundry is the Pro code door.
417
00:11:26,509 --> 00:11:29,070
It gives you control over the
model, over your own data, and
418
00:11:29,070 --> 00:11:32,029
over complex setups where
several agents work together.
419
00:11:32,029 --> 00:11:33,710
When a job needs custom logic, a
420
00:11:33,710 --> 00:11:34,950
specific model, or serious
421
00:11:34,950 --> 00:11:36,549
scale, this is the door you walk
422
00:11:36,549 --> 00:11:36,830
through.
423
00:11:36,830 --> 00:11:39,269
These are two doors into the
same house, not rivals.
424
00:11:39,269 --> 00:11:40,750
Many teams start in Copilot
425
00:11:40,750 --> 00:11:42,190
Studio for a quick win and reach
426
00:11:42,190 --> 00:11:42,669
for foundry.
427
00:11:42,669 --> 00:11:44,029
When the requirements grow.
428
00:11:44,029 --> 00:11:46,309
The two can even work side by
side, with one handling the
429
00:11:46,309 --> 00:11:48,230
conversation and the other doing
the heavy lifting.
430
00:11:48,230 --> 00:11:49,870
They also share a governance
story.
431
00:11:49,870 --> 00:11:52,190
Both sit under the same
Microsoft identity and data
432
00:11:52,190 --> 00:11:55,649
controls, so your security team
keeps one way of watching over
433
00:11:55,649 --> 00:11:56,909
agents instead of two.
434
00:11:56,909 --> 00:11:57,809
They even live on the same
435
00:11:57,809 --> 00:11:59,110
underlying platform, which is
436
00:11:59,110 --> 00:12:00,629
why the skills you learn in one
437
00:12:00,629 --> 00:12:01,590
carry straight over to the
438
00:12:01,590 --> 00:12:01,950
other.
439
00:12:01,950 --> 00:12:05,330
The practical takeaway is
simple, so start light and scale
440
00:12:05,330 --> 00:12:07,250
when you genuinely need to.
441
00:12:07,250 --> 00:12:09,049
What people actually build with
it.
442
00:12:09,049 --> 00:12:11,690
It helps to see what this looks
like inside a real company,
443
00:12:11,690 --> 00:12:14,090
because a few patterns show up
again and again.
444
00:12:14,090 --> 00:12:15,330
Customer support shows up first,
445
00:12:15,330 --> 00:12:16,490
where an agent connects to the
446
00:12:16,490 --> 00:12:17,769
system that holds your customer
447
00:12:17,769 --> 00:12:19,769
details, known as a CRM, and
448
00:12:19,769 --> 00:12:20,990
handles common questions in real
449
00:12:20,990 --> 00:12:21,570
time.
450
00:12:21,570 --> 00:12:23,169
Routine work is another where an
451
00:12:23,169 --> 00:12:24,570
agent schedules meetings, sends
452
00:12:24,570 --> 00:12:25,549
reminders, and generates
453
00:12:25,549 --> 00:12:26,009
reports.
454
00:12:26,009 --> 00:12:27,769
So people stopped doing it all
by hand.
455
00:12:27,769 --> 00:12:30,169
Then there's the quieter win,
where an agent reads current
456
00:12:30,169 --> 00:12:33,090
data and comes back with an
insight or a recommendation.
457
00:12:33,090 --> 00:12:35,649
Real examples are easier to
picture than categories.
458
00:12:35,649 --> 00:12:38,970
BMW uses multi-agent AI, so its
engineers get insights from
459
00:12:38,970 --> 00:12:41,370
vehicle measurement data that
they can act on right away.
460
00:12:41,370 --> 00:12:42,649
KPMG works in a regulated
461
00:12:42,649 --> 00:12:43,929
industry, and it leans on the
462
00:12:43,929 --> 00:12:45,250
governance and observability in
463
00:12:45,250 --> 00:12:46,529
the platform so its agents stay
464
00:12:46,529 --> 00:12:47,250
accountable.
465
00:12:47,250 --> 00:12:50,009
The same patterns turn up across
industries from a bank handling
466
00:12:50,009 --> 00:12:52,990
routine account questions to a
manufacturer pulling insight out
467
00:12:52,990 --> 00:12:54,250
of machine data.
468
00:12:54,250 --> 00:12:55,570
The details change, but the
469
00:12:55,570 --> 00:12:56,570
shape of the work stays the
470
00:12:56,570 --> 00:12:57,169
same.
471
00:12:57,169 --> 00:12:59,570
Then there's the money question,
and the answer is friendlier
472
00:12:59,570 --> 00:13:00,450
than you might expect.
473
00:13:00,450 --> 00:13:01,850
Microsoft doesn't charge extra
474
00:13:01,850 --> 00:13:03,210
for creating or running foundry
475
00:13:03,210 --> 00:13:04,450
native agents with prompts and
476
00:13:04,450 --> 00:13:05,299
workflows.
477
00:13:05,299 --> 00:13:08,259
you pay for the model usage
behind them, and for the tools
478
00:13:08,259 --> 00:13:11,299
and data you connect, which
means the cost follows the work
479
00:13:11,299 --> 00:13:13,019
the agent actually does.
480
00:13:13,019 --> 00:13:13,820
Recap.
481
00:13:13,820 --> 00:13:15,139
And one question for you.
482
00:13:15,139 --> 00:13:16,820
Let's pull this together
quickly.
483
00:13:16,820 --> 00:13:19,700
An agent is a model plus
instructions plus tools.
484
00:13:19,700 --> 00:13:22,279
And it can carry a task through
several steps rather than
485
00:13:22,279 --> 00:13:23,580
stopping at one reply.
486
00:13:23,580 --> 00:13:26,539
Microsoft Foundry Agent Service
is the platform that handles the
487
00:13:26,539 --> 00:13:29,419
parts around the model,
including the tools, the memory,
488
00:13:29,419 --> 00:13:32,340
the identity, the hosting, and
the checking of the work.
489
00:13:32,340 --> 00:13:33,779
You can start with no code and
490
00:13:33,779 --> 00:13:34,820
grow into code when you're
491
00:13:34,820 --> 00:13:35,500
ready.
492
00:13:35,500 --> 00:13:37,100
Copilot studio is the low code,
493
00:13:37,100 --> 00:13:38,759
DAW and foundry is the Pro code
494
00:13:38,759 --> 00:13:40,139
DAW and both lead into the same
495
00:13:40,139 --> 00:13:40,620
house.
496
00:13:40,620 --> 00:13:43,659
The thing worth carrying away is
this the AI brain was never the
497
00:13:43,659 --> 00:13:45,340
hard part because everything
around it was.
498
00:13:45,340 --> 00:13:47,860
And that's exactly what the
platform takes off your plate.
499
00:13:47,860 --> 00:13:49,139
So here is my question for you,
500
00:13:49,139 --> 00:13:49,980
and I'd love to hear your
501
00:13:49,980 --> 00:13:50,659
answer.
502
00:13:50,659 --> 00:13:53,679
If you could hand one task to an
agent tomorrow, what would you
503
00:13:53,679 --> 00:13:56,659
hand over first and what would
you want it to leave alone?
504
00:13:56,659 --> 00:13:58,259
If that question got you
thinking?
505
00:13:58,259 --> 00:14:00,639
Leave your answer in the
comments and subscribe to the
506
00:14:00,639 --> 00:14:03,059
M365 FM podcast for more.
507
00:14:03,059 --> 00:14:04,740
Simply explained episodes that
508
00:14:04,740 --> 00:14:05,820
make this world a little less
509
00:14:05,820 --> 00:14:06,820
intimidating.
00:00:00,040 --> 00:00:01,600
Microsoft's Foundry catalog now
2
00:00:01,600 --> 00:00:03,020
holds more than ten thousand AI
3
00:00:03,020 --> 00:00:03,720
models.
4
00:00:03,720 --> 00:00:06,080
That number only matters if you
can turn a model into something
5
00:00:06,080 --> 00:00:07,160
that actually does work.
6
00:00:07,160 --> 00:00:08,160
You might assume that means
7
00:00:08,160 --> 00:00:09,800
hiring developers and setting up
8
00:00:09,800 --> 00:00:10,439
servers.
9
00:00:10,439 --> 00:00:12,060
Foundry agent service is how
10
00:00:12,060 --> 00:00:13,599
Microsoft lets you build those
11
00:00:13,599 --> 00:00:15,240
working agents without the
12
00:00:15,240 --> 00:00:16,719
server headache.
13
00:00:16,719 --> 00:00:20,120
Welcome to the M365 FM podcast,
where we explore the people,
14
00:00:20,120 --> 00:00:22,600
ideas, and technologies shaping
the future of work.
15
00:00:22,600 --> 00:00:24,679
By the end, you'll know what it
is, how it works, and whether
16
00:00:24,679 --> 00:00:26,039
you could use it.
17
00:00:26,039 --> 00:00:27,960
What an AI agent actually is.
18
00:00:27,960 --> 00:00:30,440
Start with the thing you already
know because it makes the rest
19
00:00:30,440 --> 00:00:31,440
of this easier.
20
00:00:31,440 --> 00:00:33,920
A chatbot waits for your
question, gives you an answer,
21
00:00:33,920 --> 00:00:35,439
and then stops right there.
22
00:00:35,439 --> 00:00:36,719
An agent takes your request and
23
00:00:36,719 --> 00:00:37,840
keeps working so it can look
24
00:00:37,840 --> 00:00:38,439
something up.
25
00:00:38,439 --> 00:00:40,719
Change a record, and then tell
you what it did.
26
00:00:40,719 --> 00:00:43,439
Picture a helper who handles
your travel booking for you.
27
00:00:43,439 --> 00:00:45,840
You say you need a trip next
month and the helper checks your
28
00:00:45,840 --> 00:00:49,119
calendar, finds a flight, books
it, and sends you the details.
29
00:00:49,119 --> 00:00:51,359
A chatbot would only describe
how to book that flight while an
30
00:00:51,359 --> 00:00:52,719
agent goes ahead and books it.
31
00:00:52,719 --> 00:00:55,479
Plane automation is different
again, and it's worth keeping
32
00:00:55,479 --> 00:00:56,439
the two apart.
33
00:00:56,439 --> 00:00:57,920
Automation follows a fixed set
34
00:00:57,920 --> 00:00:59,439
of steps every single time, so
35
00:00:59,439 --> 00:01:00,560
it breaks the moment something
36
00:01:00,560 --> 00:01:01,979
unexpected happens.
37
00:01:01,979 --> 00:01:05,060
An agent decides which step to
take next based on whatever it
38
00:01:05,060 --> 00:01:07,659
finds along the way, which is
why it holds up better when the
39
00:01:07,659 --> 00:01:08,859
real world gets messy.
40
00:01:08,859 --> 00:01:11,420
If that sounds like a small
difference, think about how much
41
00:01:11,420 --> 00:01:13,739
of your day disappears into
chasing things down.
42
00:01:13,739 --> 00:01:14,980
Someone asks for an update and
43
00:01:14,980 --> 00:01:16,719
you open three systems to
44
00:01:16,719 --> 00:01:17,340
answer.
45
00:01:17,340 --> 00:01:18,859
An agent can do that opening and
46
00:01:18,859 --> 00:01:20,299
checking for you, and then hand
47
00:01:20,299 --> 00:01:21,260
you the answer with the work
48
00:01:21,260 --> 00:01:22,099
already done.
49
00:01:22,099 --> 00:01:24,019
Every agent needs three simple
parts.
50
00:01:24,019 --> 00:01:25,420
The first part is a model, which
51
00:01:25,420 --> 00:01:26,540
is the AI that reads your
52
00:01:26,540 --> 00:01:27,780
request and decides what to do
53
00:01:27,780 --> 00:01:28,459
next.
54
00:01:28,459 --> 00:01:29,900
The second part is instructions,
55
00:01:29,900 --> 00:01:31,140
which act like a job description
56
00:01:31,140 --> 00:01:32,299
and tell the agent what it is
57
00:01:32,299 --> 00:01:32,900
for.
58
00:01:32,900 --> 00:01:35,900
The third part is tools, which
are the ways the agent can act,
59
00:01:35,900 --> 00:01:38,099
such as searching the web or
updating a record.
60
00:01:38,099 --> 00:01:39,420
Put those three parts together
61
00:01:39,420 --> 00:01:40,659
and you get something that can
62
00:01:40,659 --> 00:01:42,019
work through a task in several
63
00:01:42,019 --> 00:01:43,299
steps instead of stopping at one
64
00:01:43,299 --> 00:01:44,019
reply.
65
00:01:44,019 --> 00:01:45,260
That's the whole idea behind
66
00:01:45,260 --> 00:01:46,700
every agent you'll meet today,
67
00:01:46,700 --> 00:01:47,780
no matter how big the name
68
00:01:47,780 --> 00:01:48,500
sounds.
69
00:01:48,500 --> 00:01:50,180
The model, it turns out, is the
easy part.
70
00:01:50,180 --> 00:01:51,379
Everything around it is where
71
00:01:51,379 --> 00:01:52,939
the real work lives, and that
72
00:01:52,939 --> 00:01:54,099
idea comes back in a few
73
00:01:54,099 --> 00:01:55,180
minutes.
74
00:01:55,180 --> 00:01:57,299
What Microsoft Founder Agent
service is.
75
00:01:57,299 --> 00:01:58,900
Now let's put a name on the
product.
76
00:01:58,900 --> 00:02:01,704
Microsoft Foundry is the
platform, and Foundry agent
77
00:02:01,704 --> 00:02:04,430
service is the part of it built
for creating and running agents.
78
00:02:04,430 --> 00:02:05,670
It lives in Azure, which is
79
00:02:05,670 --> 00:02:07,150
simply Microsoft's cloud, the
80
00:02:07,150 --> 00:02:08,189
place where companies rent
81
00:02:08,189 --> 00:02:09,310
computing power instead of
82
00:02:09,310 --> 00:02:10,469
buying and maintaining their own
83
00:02:10,469 --> 00:02:11,270
machines.
84
00:02:11,270 --> 00:02:13,189
The name is a small clue as
well.
85
00:02:13,189 --> 00:02:16,189
A foundry is a place where raw
material goes in and finished.
86
00:02:16,189 --> 00:02:18,710
Things come out, and that's
roughly the idea here.
87
00:02:18,710 --> 00:02:19,990
Models and tools go in and
88
00:02:19,990 --> 00:02:21,069
working agents come out the
89
00:02:21,069 --> 00:02:21,830
other side.
90
00:02:21,830 --> 00:02:23,669
The word that matters most here
is managed.
91
00:02:23,669 --> 00:02:24,830
When something is managed,
92
00:02:24,830 --> 00:02:26,189
Microsoft looks after the
93
00:02:26,189 --> 00:02:27,409
servers, the scaling and the
94
00:02:27,409 --> 00:02:28,110
updates.
95
00:02:28,110 --> 00:02:29,389
So you spend your time on the
96
00:02:29,389 --> 00:02:30,430
agent instead of on the
97
00:02:30,430 --> 00:02:30,990
plumbing.
98
00:02:30,990 --> 00:02:32,189
You still pay for what you use,
99
00:02:32,189 --> 00:02:33,370
but you don't sit up at two in
100
00:02:33,370 --> 00:02:34,349
the morning restarting a
101
00:02:34,349 --> 00:02:34,830
machine.
102
00:02:34,830 --> 00:02:37,349
That single word explains why a
beginner can get something.
103
00:02:37,349 --> 00:02:40,310
Working in an afternoon foundry
also meets you at whatever level
104
00:02:40,310 --> 00:02:41,389
you're comfortable with.
105
00:02:41,389 --> 00:02:44,750
At one end, you build an agent
with nothing but settings and
106
00:02:44,750 --> 00:02:47,909
plain written instructions,
which means no coding at all.
107
00:02:47,909 --> 00:02:50,710
At the other end, you bring your
own code and your own framework
108
00:02:50,710 --> 00:02:52,310
and foundry runs it for you.
109
00:02:52,310 --> 00:02:53,990
You can start at the simple end
110
00:02:53,990 --> 00:02:55,310
and move toward the code end
111
00:02:55,310 --> 00:02:56,909
later without throwing away what
112
00:02:56,909 --> 00:02:57,629
you built.
113
00:02:57,629 --> 00:02:59,870
There are three kinds of agent
to know about, and the names are
114
00:02:59,870 --> 00:03:01,189
friendlier than they look.
115
00:03:01,189 --> 00:03:04,169
A prompt agent is built from
settings and instructions, and
116
00:03:04,169 --> 00:03:06,090
it's the fastest place to begin.
117
00:03:06,090 --> 00:03:07,210
A voice agent is built for
118
00:03:07,210 --> 00:03:08,810
spoken conversations, so it can
119
00:03:08,810 --> 00:03:09,969
answer on a phone line or in a
120
00:03:09,969 --> 00:03:10,689
browser.
121
00:03:10,689 --> 00:03:13,770
A hosted agent is your own code,
packaged up and run by foundry,
122
00:03:13,770 --> 00:03:16,009
with the servers and the scaling
handled for you.
123
00:03:16,009 --> 00:03:18,330
Most people start with the first
kind and never need the other
124
00:03:18,330 --> 00:03:19,889
two, which is perfectly fine.
125
00:03:19,889 --> 00:03:21,289
Underneath all three, foundry
126
00:03:21,289 --> 00:03:23,169
brings agents, models, and tools
127
00:03:23,169 --> 00:03:24,650
together in one place with
128
00:03:24,650 --> 00:03:25,969
shared access control and shared
129
00:03:25,969 --> 00:03:26,849
monitoring.
130
00:03:26,849 --> 00:03:28,289
In practice, that means one
131
00:03:28,289 --> 00:03:29,629
place to sign in, one place to
132
00:03:29,629 --> 00:03:31,090
set permissions, and one place
133
00:03:31,090 --> 00:03:32,250
to watch what your agents are
134
00:03:32,250 --> 00:03:32,689
doing.
135
00:03:32,689 --> 00:03:33,210
That matters.
136
00:03:33,210 --> 00:03:36,169
Later, when you want to know who
can touch an agent and what that
137
00:03:36,169 --> 00:03:37,650
agent has been doing.
138
00:03:37,650 --> 00:03:38,729
Models the brain.
139
00:03:38,729 --> 00:03:39,610
You can swap.
140
00:03:39,610 --> 00:03:41,169
Every agent runs on a model, and
141
00:03:41,169 --> 00:03:42,610
the model is the part that does
142
00:03:42,610 --> 00:03:43,330
the thinking.
143
00:03:43,330 --> 00:03:44,689
Think of it as the brain of the
144
00:03:44,689 --> 00:03:46,550
agent because it decides what to
145
00:03:46,550 --> 00:03:48,650
say or do while the tools handle
146
00:03:48,650 --> 00:03:49,370
the acting.
147
00:03:49,370 --> 00:03:51,610
The scale here is bigger than
most people expect.
148
00:03:51,610 --> 00:03:54,610
The foundry catalog offers more
than ten thousand models built
149
00:03:54,610 --> 00:03:58,090
by Microsoft, OpenAI, anthropic,
meta, and plenty of others.
150
00:03:58,090 --> 00:03:59,569
Having that many options sounds
151
00:03:59,569 --> 00:04:01,129
overwhelming, but you don't have
152
00:04:01,129 --> 00:04:02,280
to shop around for long, because
153
00:04:02,280 --> 00:04:03,509
you can start with a sensible
154
00:04:03,509 --> 00:04:04,789
default and change your mind
155
00:04:04,789 --> 00:04:05,349
later.
156
00:04:05,349 --> 00:04:08,030
The useful part for you is that
you can swap the model without
157
00:04:08,030 --> 00:04:09,189
rewriting your agent.
158
00:04:09,189 --> 00:04:12,069
If a cheap, fast model does the
job, you use that one.
159
00:04:12,069 --> 00:04:14,550
If a harder task needs a
stronger model, you change a
160
00:04:14,550 --> 00:04:16,949
single setting and keep
everything else exactly as it
161
00:04:16,949 --> 00:04:19,829
was that freedom matters because
it means you aren't locked into
162
00:04:19,829 --> 00:04:21,110
a choice you made on day one.
163
00:04:21,110 --> 00:04:22,350
Different jobs suit different
164
00:04:22,350 --> 00:04:23,350
models in the same way that
165
00:04:23,350 --> 00:04:24,910
different errands suit different
166
00:04:24,910 --> 00:04:25,790
vehicles.
167
00:04:25,790 --> 00:04:28,269
A quick reply is a small errand,
and a long chain of careful
168
00:04:28,269 --> 00:04:29,509
reasoning is a big one.
169
00:04:29,509 --> 00:04:30,709
The model you pick should match
170
00:04:30,709 --> 00:04:31,870
the weight of the task in front
171
00:04:31,870 --> 00:04:32,350
of it.
172
00:04:32,350 --> 00:04:35,149
The catalog also spans a wide
range, from tiny models that
173
00:04:35,149 --> 00:04:37,870
cost almost nothing to large
ones that reason through hard
174
00:04:37,870 --> 00:04:39,509
problems step by step.
175
00:04:39,509 --> 00:04:41,870
You don't need to memorize any
of it, because the point is
176
00:04:41,870 --> 00:04:44,509
simply that you have options and
you can change your mind without
177
00:04:44,509 --> 00:04:45,829
rebuilding the agent.
178
00:04:45,829 --> 00:04:48,189
So you have a brain you can
swap, and that's genuinely
179
00:04:48,189 --> 00:04:49,069
useful on its own.
180
00:04:49,069 --> 00:04:51,670
The trouble is that a brain
can't lift a finger by itself,
181
00:04:51,670 --> 00:04:54,550
and that gap is exactly what the
next part fills.
182
00:04:54,550 --> 00:04:55,670
Tools and knowledge.
183
00:04:55,670 --> 00:04:57,310
How agents take action.
184
00:04:57,310 --> 00:04:58,790
A model on its own can only
185
00:04:58,790 --> 00:05:00,430
talk, and talking isn't the same
186
00:05:00,430 --> 00:05:01,189
as doing.
187
00:05:01,189 --> 00:05:03,129
To do real work, an agent needs
188
00:05:03,129 --> 00:05:04,730
tools and a tool is simply a
189
00:05:04,730 --> 00:05:05,970
capability the agent is allowed
190
00:05:05,970 --> 00:05:06,370
to use.
191
00:05:06,370 --> 00:05:09,250
If the model is the brain, the
tools are the hands.
192
00:05:09,250 --> 00:05:12,009
Foundry ships with built in
tools that you can switch on.
193
00:05:12,009 --> 00:05:13,569
Web search looks things up.
194
00:05:13,569 --> 00:05:16,730
File search reads your documents
and a code interpreter runs
195
00:05:16,730 --> 00:05:18,370
calculations and small programs.
196
00:05:18,370 --> 00:05:21,449
Memory is the forth and it lets
the agent hold on to useful
197
00:05:21,449 --> 00:05:22,769
details between conversations.
198
00:05:22,769 --> 00:05:25,290
These all come ready to use so
you get real value before you
199
00:05:25,290 --> 00:05:26,529
connect anything of your own.
200
00:05:26,529 --> 00:05:27,610
You aren't limited to those
201
00:05:27,610 --> 00:05:28,689
because you can add your own
202
00:05:28,689 --> 00:05:29,610
tools too.
203
00:05:29,610 --> 00:05:30,889
You can connect a tool through a
204
00:05:30,889 --> 00:05:32,389
function, through an open API
205
00:05:32,389 --> 00:05:34,689
definition or through MCP, which
206
00:05:34,689 --> 00:05:36,290
is a shared plug standard that
207
00:05:36,290 --> 00:05:38,050
lets AI tools connect to other
208
00:05:38,050 --> 00:05:38,889
systems.
209
00:05:38,889 --> 00:05:41,810
The exact labels matter less
than the idea, which is that
210
00:05:41,810 --> 00:05:44,889
your agent can reach the systems
your business already runs on.
211
00:05:44,889 --> 00:05:46,529
Knowledge is the other half of
the story.
212
00:05:46,529 --> 00:05:47,649
You can point an agent at your
213
00:05:47,649 --> 00:05:49,290
own content, such as files in
214
00:05:49,290 --> 00:05:50,930
SharePoint, data in Microsoft
215
00:05:50,930 --> 00:05:52,689
Fabric, or current results from
216
00:05:52,689 --> 00:05:53,490
Bing Search.
217
00:05:53,490 --> 00:05:56,490
When an agent answers from your
real data instead of guessing.
218
00:05:56,490 --> 00:05:57,410
That's called grounding.
219
00:05:57,410 --> 00:05:58,490
And it's the difference between
220
00:05:58,490 --> 00:06:00,209
a confident answer and a correct
221
00:06:00,209 --> 00:06:00,730
one.
222
00:06:00,730 --> 00:06:03,500
One more piece is worth knowing,
and it's called a toolbox.
223
00:06:03,500 --> 00:06:04,779
A toolbox lets you gather your
224
00:06:04,779 --> 00:06:06,579
tools once, expose them through
225
00:06:06,579 --> 00:06:08,259
a single connection, and reuse
226
00:06:08,259 --> 00:06:09,939
them across many agents with
227
00:06:09,939 --> 00:06:10,980
shared sign in and shared
228
00:06:10,980 --> 00:06:11,779
controls.
229
00:06:11,779 --> 00:06:12,980
It saves you from wiring the
230
00:06:12,980 --> 00:06:14,420
same tool into every new agent
231
00:06:14,420 --> 00:06:15,100
you build.
232
00:06:15,100 --> 00:06:16,459
To make this concrete picture an
233
00:06:16,459 --> 00:06:17,620
agent that helps onboard a new
234
00:06:17,620 --> 00:06:18,420
employee.
235
00:06:18,420 --> 00:06:19,860
In one request, it could pull
236
00:06:19,860 --> 00:06:21,120
the onboarding guidance from a
237
00:06:21,120 --> 00:06:22,420
knowledge base, create the
238
00:06:22,420 --> 00:06:24,019
account, draft a welcome email,
239
00:06:24,019 --> 00:06:25,500
and post a message in the team's
240
00:06:25,500 --> 00:06:26,180
channel.
241
00:06:26,180 --> 00:06:28,980
That's several tools and several
systems working together behind
242
00:06:28,980 --> 00:06:30,019
a single request.
243
00:06:30,019 --> 00:06:32,620
And it's the moment an agent
stops being a novelty and starts
244
00:06:32,620 --> 00:06:34,060
being genuinely useful.
245
00:06:34,060 --> 00:06:36,220
Now you can see why the model
was the easy part.
246
00:06:36,220 --> 00:06:37,500
The work sits in the tools, the
247
00:06:37,500 --> 00:06:38,939
data, and the permissions around
248
00:06:38,939 --> 00:06:40,300
them, and a toolbox is
249
00:06:40,300 --> 00:06:41,699
Microsoft's way of making that
250
00:06:41,699 --> 00:06:41,980
work.
251
00:06:41,980 --> 00:06:46,019
Manageable threads, runs,
conversations and responses.
252
00:06:46,019 --> 00:06:48,980
Memory is what makes an agent
feel less like a search box and
253
00:06:48,980 --> 00:06:51,699
more like a colleague, and it
comes down to a few terms that
254
00:06:51,699 --> 00:06:53,139
used to confuse everyone.
255
00:06:53,139 --> 00:06:54,139
The good news is that those
256
00:06:54,139 --> 00:06:55,180
terms are getting simpler, so
257
00:06:55,180 --> 00:06:56,339
this part is easier than it
258
00:06:56,339 --> 00:06:57,180
sounds.
259
00:06:57,180 --> 00:06:58,660
In the older way of doing
things.
260
00:06:58,660 --> 00:07:00,420
A conversation was called a
thread.
261
00:07:00,420 --> 00:07:03,149
A thread held your messages,
which were the back and forth
262
00:07:03,149 --> 00:07:04,360
between you and the agent.
263
00:07:04,360 --> 00:07:07,079
Each time the agent took a turn,
that was called a run, and you
264
00:07:07,079 --> 00:07:09,720
could watch the run move from
waiting to working to finished.
265
00:07:09,720 --> 00:07:12,519
The new away keeps the same idea
and changes the names.
266
00:07:12,519 --> 00:07:15,240
A conversation now holds the
history and a response is what
267
00:07:15,240 --> 00:07:18,000
the agent produces, which can be
an answer or an action.
268
00:07:18,000 --> 00:07:21,800
So threads became conversations
and runs became responses.
269
00:07:21,800 --> 00:07:23,959
And that's the whole rename in a
single line.
270
00:07:23,959 --> 00:07:25,879
The change does more than tidy
up the wording.
271
00:07:25,879 --> 00:07:27,160
A conversation can store more
272
00:07:27,160 --> 00:07:28,600
than messages because it can
273
00:07:28,600 --> 00:07:30,120
also keep track of the tools the
274
00:07:30,120 --> 00:07:31,240
agent called and what those
275
00:07:31,240 --> 00:07:32,240
tools returned.
276
00:07:32,240 --> 00:07:34,680
The agent ends up with a richer
memory of what happened, which
277
00:07:34,680 --> 00:07:36,639
makes it better at picking up
where you left off.
278
00:07:36,639 --> 00:07:38,199
History matters for a practical
reason.
279
00:07:38,199 --> 00:07:39,800
Most useful tasks take more than
280
00:07:39,800 --> 00:07:41,439
one message, and an agent that
281
00:07:41,439 --> 00:07:42,639
forgets the first half of a
282
00:07:42,639 --> 00:07:43,899
request isn't much use to
283
00:07:43,899 --> 00:07:44,519
anyone.
284
00:07:44,519 --> 00:07:47,360
Keeping the conversation lets
the agent follow a thread of
285
00:07:47,360 --> 00:07:49,720
thought across many turns the
way a person would.
286
00:07:49,720 --> 00:07:52,759
You felt this in every good chat
tool you've used, you ask a
287
00:07:52,759 --> 00:07:55,360
follow up question without
repeating yourself, and the tool
288
00:07:55,360 --> 00:07:58,160
understands because it kept the
earlier turns in view.
289
00:07:58,160 --> 00:07:59,439
Conversations give an agent that
290
00:07:59,439 --> 00:08:00,740
same ability, which is why they
291
00:08:00,740 --> 00:08:01,800
sit at the center of how it
292
00:08:01,800 --> 00:08:02,639
works.
293
00:08:02,639 --> 00:08:03,899
One practical note, since you
294
00:08:03,899 --> 00:08:05,660
may have read older guides, the
295
00:08:05,660 --> 00:08:07,060
old threads and runs setup is
296
00:08:07,060 --> 00:08:08,300
being retired in favor of this
297
00:08:08,300 --> 00:08:08,980
newer one.
298
00:08:08,980 --> 00:08:10,160
So if you're learning today,
299
00:08:10,160 --> 00:08:11,420
learn the words, conversation
300
00:08:11,420 --> 00:08:12,339
and response.
301
00:08:12,339 --> 00:08:14,180
That's where the platform has
moved.
302
00:08:14,180 --> 00:08:15,379
Hosting and scaling.
303
00:08:15,379 --> 00:08:16,660
Where the agent lives.
304
00:08:16,660 --> 00:08:17,980
An agent has to run somewhere
305
00:08:17,980 --> 00:08:19,220
and that somewhere has to stay
306
00:08:19,220 --> 00:08:20,620
available when people actually
307
00:08:20,620 --> 00:08:20,980
use it.
308
00:08:20,980 --> 00:08:21,819
That's what hosting means.
309
00:08:21,819 --> 00:08:24,939
And it's one of those jobs you'd
much rather hand to someone else
310
00:08:24,939 --> 00:08:25,540
for a prompt.
311
00:08:25,540 --> 00:08:27,379
Agent hosting is barely a topic
at all.
312
00:08:27,379 --> 00:08:30,000
Microsoft runs it for you, so
there's nothing to install,
313
00:08:30,000 --> 00:08:31,060
patch, or restart.
314
00:08:31,060 --> 00:08:33,299
You write the instructions,
choose the tools, and the agent
315
00:08:33,299 --> 00:08:35,820
is simply there when someone
calls it hosted agents need a
316
00:08:35,820 --> 00:08:38,820
little more explanation because
you're bringing your own code.
317
00:08:38,820 --> 00:08:40,340
Foundry runs that code in its
318
00:08:40,340 --> 00:08:41,740
own isolated space for each
319
00:08:41,740 --> 00:08:43,460
session, which is a tidy way of
320
00:08:43,460 --> 00:08:44,899
saying that one person's work
321
00:08:44,899 --> 00:08:46,059
stays separate from everyone
322
00:08:46,059 --> 00:08:46,740
else's.
323
00:08:46,740 --> 00:08:48,059
When more people use the agent,
324
00:08:48,059 --> 00:08:49,539
it scales up on its own, and
325
00:08:49,539 --> 00:08:51,179
when fewer do, it scales back
326
00:08:51,179 --> 00:08:51,700
down.
327
00:08:51,700 --> 00:08:54,460
The idle behavior is worth
pausing on because it answers
328
00:08:54,460 --> 00:08:55,820
the cost question.
329
00:08:55,820 --> 00:08:58,779
When nobody is using an agent,
it goes to sleep and its state
330
00:08:58,779 --> 00:08:59,740
is kept safe.
331
00:08:59,740 --> 00:09:02,620
When the next person arrives, it
wakes up exactly where it left
332
00:09:02,620 --> 00:09:05,950
off, and you aren't paying for
the quiet hours in between.
333
00:09:05,950 --> 00:09:09,190
Every hosted agent also gets its
own identity and its own address
334
00:09:09,190 --> 00:09:11,789
automatically, and that sets up
the next part of the story.
335
00:09:11,789 --> 00:09:14,950
That identity is how the agent
proves who it is when it reaches
336
00:09:14,950 --> 00:09:15,909
for your systems.
337
00:09:15,909 --> 00:09:18,190
One more detail is worth
knowing, even if it stays
338
00:09:18,190 --> 00:09:19,590
invisible to you.
339
00:09:19,590 --> 00:09:22,350
Hosted agents run in a long list
of Azure regions around the
340
00:09:22,350 --> 00:09:24,909
world, so a company can often
keep its agents close to where
341
00:09:24,909 --> 00:09:26,509
its data already lives.
342
00:09:26,509 --> 00:09:28,750
For the audience here, the short
version is that scaling is a
343
00:09:28,750 --> 00:09:31,230
problem the platform has already
solved for you.
344
00:09:31,230 --> 00:09:33,269
Safety, identity and evaluation.
345
00:09:33,269 --> 00:09:34,389
Once an agent can reach into
346
00:09:34,389 --> 00:09:35,950
your systems, trust becomes the
347
00:09:35,950 --> 00:09:36,629
whole game.
348
00:09:36,629 --> 00:09:39,350
And this is where Managed
platform earns its place.
349
00:09:39,350 --> 00:09:40,549
Start with identity.
350
00:09:40,549 --> 00:09:42,309
Every agent can get its own sign
351
00:09:42,309 --> 00:09:43,669
in, which means it only touches
352
00:09:43,669 --> 00:09:44,909
the things you've allowed and
353
00:09:44,909 --> 00:09:45,750
nothing else.
354
00:09:45,750 --> 00:09:47,169
You can also see what it did so
355
00:09:47,169 --> 00:09:48,789
an agent stays accountable in a
356
00:09:48,789 --> 00:09:50,110
way a shared password never
357
00:09:50,110 --> 00:09:50,669
could.
358
00:09:50,669 --> 00:09:52,029
Guardrails are the second layer.
359
00:09:52,029 --> 00:09:54,909
These are built in filters that
help block unsafe output.
360
00:09:54,909 --> 00:09:56,110
They also reduce the risk of
361
00:09:56,110 --> 00:09:57,190
someone tricking the agent
362
00:09:57,190 --> 00:09:58,470
through a prompt, which is the
363
00:09:58,470 --> 00:09:59,909
practice of slipping hidden
364
00:09:59,909 --> 00:10:01,230
instructions into something the
365
00:10:01,230 --> 00:10:02,230
agent reads.
366
00:10:02,230 --> 00:10:04,889
You can also keep an agent
inside your own private network
367
00:10:04,889 --> 00:10:07,370
so its data doesn't wander
anywhere you haven't approved.
368
00:10:07,370 --> 00:10:09,610
Evaluation is the part that
surprises people because it
369
00:10:09,610 --> 00:10:11,009
works a lot like testing.
370
00:10:11,009 --> 00:10:12,610
Foundry includes checkers that
score.
371
00:10:12,610 --> 00:10:13,690
Whether the agent finished the
372
00:10:13,690 --> 00:10:15,129
task, whether it followed your
373
00:10:15,129 --> 00:10:16,610
instructions, and whether it
374
00:10:16,610 --> 00:10:18,009
used the right tools along the
375
00:10:18,009 --> 00:10:18,690
way.
376
00:10:18,690 --> 00:10:21,490
You can think of them as unit
tests for an agent, and they run
377
00:10:21,490 --> 00:10:23,370
before real people have to
depend on it.
378
00:10:23,370 --> 00:10:27,090
Observability is the last piece,
and it simply means you can
379
00:10:27,090 --> 00:10:28,850
trace every step the agent took.
380
00:10:28,850 --> 00:10:29,889
When something goes wrong, you
381
00:10:29,889 --> 00:10:30,909
follow the trail instead of
382
00:10:30,909 --> 00:10:31,250
guessing.
383
00:10:31,250 --> 00:10:33,129
And that's how confidence builds
over time.
384
00:10:33,129 --> 00:10:35,490
None of this means you can
switch off your own judgment.
385
00:10:35,490 --> 00:10:38,090
A managed platform gives you the
controls, but you still decide
386
00:10:38,090 --> 00:10:41,169
what an agent is allowed to do
and how far it's trusted.
387
00:10:41,169 --> 00:10:42,389
The sensible approach is to
388
00:10:42,389 --> 00:10:43,750
start with a small, low risk
389
00:10:43,750 --> 00:10:44,490
task.
390
00:10:44,490 --> 00:10:45,929
Watch how the agent behaves and
391
00:10:45,929 --> 00:10:47,490
widen its reach only as your
392
00:10:47,490 --> 00:10:48,409
confidence grows.
393
00:10:48,409 --> 00:10:50,970
When an agent is ready, you can
publish it to Microsoft Teams
394
00:10:50,970 --> 00:10:53,090
and Microsoft three hundred
sixty five copilot, so people
395
00:10:53,090 --> 00:10:55,289
meet it in the tools they
already use every day.
396
00:10:55,289 --> 00:10:57,730
That's what separates a clever
experiment from something your
397
00:10:57,730 --> 00:10:59,250
team actually relies on.
398
00:10:59,250 --> 00:11:01,210
How it relates to Copilot
Studio.
399
00:11:01,210 --> 00:11:02,450
By now, a fair question is
400
00:11:02,450 --> 00:11:03,889
forming because Microsoft has
401
00:11:03,889 --> 00:11:05,110
more than one way to build an
402
00:11:05,110 --> 00:11:06,529
agent and the names start to
403
00:11:06,529 --> 00:11:07,470
blur together.
404
00:11:07,470 --> 00:11:08,309
Let's clear that up.
405
00:11:08,309 --> 00:11:10,389
Copilot studio is the low code
door.
406
00:11:10,389 --> 00:11:11,629
It's built for business users
407
00:11:11,629 --> 00:11:12,909
and makers, and it connects
408
00:11:12,909 --> 00:11:14,149
quickly to the Microsoft three
409
00:11:14,149 --> 00:11:15,309
hundred sixty five tools and
410
00:11:15,309 --> 00:11:16,590
power platform connectors your
411
00:11:16,590 --> 00:11:18,269
company already owns for many
412
00:11:18,269 --> 00:11:18,789
teams.
413
00:11:18,789 --> 00:11:20,990
It's the fastest route to a
working agent.
414
00:11:20,990 --> 00:11:22,429
If your goal is a helpful
assistant.
415
00:11:22,429 --> 00:11:24,870
Inside teams, this is usually
where you start.
416
00:11:24,870 --> 00:11:26,509
Foundry is the Pro code door.
417
00:11:26,509 --> 00:11:29,070
It gives you control over the
model, over your own data, and
418
00:11:29,070 --> 00:11:32,029
over complex setups where
several agents work together.
419
00:11:32,029 --> 00:11:33,710
When a job needs custom logic, a
420
00:11:33,710 --> 00:11:34,950
specific model, or serious
421
00:11:34,950 --> 00:11:36,549
scale, this is the door you walk
422
00:11:36,549 --> 00:11:36,830
through.
423
00:11:36,830 --> 00:11:39,269
These are two doors into the
same house, not rivals.
424
00:11:39,269 --> 00:11:40,750
Many teams start in Copilot
425
00:11:40,750 --> 00:11:42,190
Studio for a quick win and reach
426
00:11:42,190 --> 00:11:42,669
for foundry.
427
00:11:42,669 --> 00:11:44,029
When the requirements grow.
428
00:11:44,029 --> 00:11:46,309
The two can even work side by
side, with one handling the
429
00:11:46,309 --> 00:11:48,230
conversation and the other doing
the heavy lifting.
430
00:11:48,230 --> 00:11:49,870
They also share a governance
story.
431
00:11:49,870 --> 00:11:52,190
Both sit under the same
Microsoft identity and data
432
00:11:52,190 --> 00:11:55,649
controls, so your security team
keeps one way of watching over
433
00:11:55,649 --> 00:11:56,909
agents instead of two.
434
00:11:56,909 --> 00:11:57,809
They even live on the same
435
00:11:57,809 --> 00:11:59,110
underlying platform, which is
436
00:11:59,110 --> 00:12:00,629
why the skills you learn in one
437
00:12:00,629 --> 00:12:01,590
carry straight over to the
438
00:12:01,590 --> 00:12:01,950
other.
439
00:12:01,950 --> 00:12:05,330
The practical takeaway is
simple, so start light and scale
440
00:12:05,330 --> 00:12:07,250
when you genuinely need to.
441
00:12:07,250 --> 00:12:09,049
What people actually build with
it.
442
00:12:09,049 --> 00:12:11,690
It helps to see what this looks
like inside a real company,
443
00:12:11,690 --> 00:12:14,090
because a few patterns show up
again and again.
444
00:12:14,090 --> 00:12:15,330
Customer support shows up first,
445
00:12:15,330 --> 00:12:16,490
where an agent connects to the
446
00:12:16,490 --> 00:12:17,769
system that holds your customer
447
00:12:17,769 --> 00:12:19,769
details, known as a CRM, and
448
00:12:19,769 --> 00:12:20,990
handles common questions in real
449
00:12:20,990 --> 00:12:21,570
time.
450
00:12:21,570 --> 00:12:23,169
Routine work is another where an
451
00:12:23,169 --> 00:12:24,570
agent schedules meetings, sends
452
00:12:24,570 --> 00:12:25,549
reminders, and generates
453
00:12:25,549 --> 00:12:26,009
reports.
454
00:12:26,009 --> 00:12:27,769
So people stopped doing it all
by hand.
455
00:12:27,769 --> 00:12:30,169
Then there's the quieter win,
where an agent reads current
456
00:12:30,169 --> 00:12:33,090
data and comes back with an
insight or a recommendation.
457
00:12:33,090 --> 00:12:35,649
Real examples are easier to
picture than categories.
458
00:12:35,649 --> 00:12:38,970
BMW uses multi-agent AI, so its
engineers get insights from
459
00:12:38,970 --> 00:12:41,370
vehicle measurement data that
they can act on right away.
460
00:12:41,370 --> 00:12:42,649
KPMG works in a regulated
461
00:12:42,649 --> 00:12:43,929
industry, and it leans on the
462
00:12:43,929 --> 00:12:45,250
governance and observability in
463
00:12:45,250 --> 00:12:46,529
the platform so its agents stay
464
00:12:46,529 --> 00:12:47,250
accountable.
465
00:12:47,250 --> 00:12:50,009
The same patterns turn up across
industries from a bank handling
466
00:12:50,009 --> 00:12:52,990
routine account questions to a
manufacturer pulling insight out
467
00:12:52,990 --> 00:12:54,250
of machine data.
468
00:12:54,250 --> 00:12:55,570
The details change, but the
469
00:12:55,570 --> 00:12:56,570
shape of the work stays the
470
00:12:56,570 --> 00:12:57,169
same.
471
00:12:57,169 --> 00:12:59,570
Then there's the money question,
and the answer is friendlier
472
00:12:59,570 --> 00:13:00,450
than you might expect.
473
00:13:00,450 --> 00:13:01,850
Microsoft doesn't charge extra
474
00:13:01,850 --> 00:13:03,210
for creating or running foundry
475
00:13:03,210 --> 00:13:04,450
native agents with prompts and
476
00:13:04,450 --> 00:13:05,299
workflows.
477
00:13:05,299 --> 00:13:08,259
you pay for the model usage
behind them, and for the tools
478
00:13:08,259 --> 00:13:11,299
and data you connect, which
means the cost follows the work
479
00:13:11,299 --> 00:13:13,019
the agent actually does.
480
00:13:13,019 --> 00:13:13,820
Recap.
481
00:13:13,820 --> 00:13:15,139
And one question for you.
482
00:13:15,139 --> 00:13:16,820
Let's pull this together
quickly.
483
00:13:16,820 --> 00:13:19,700
An agent is a model plus
instructions plus tools.
484
00:13:19,700 --> 00:13:22,279
And it can carry a task through
several steps rather than
485
00:13:22,279 --> 00:13:23,580
stopping at one reply.
486
00:13:23,580 --> 00:13:26,539
Microsoft Foundry Agent Service
is the platform that handles the
487
00:13:26,539 --> 00:13:29,419
parts around the model,
including the tools, the memory,
488
00:13:29,419 --> 00:13:32,340
the identity, the hosting, and
the checking of the work.
489
00:13:32,340 --> 00:13:33,779
You can start with no code and
490
00:13:33,779 --> 00:13:34,820
grow into code when you're
491
00:13:34,820 --> 00:13:35,500
ready.
492
00:13:35,500 --> 00:13:37,100
Copilot studio is the low code,
493
00:13:37,100 --> 00:13:38,759
DAW and foundry is the Pro code
494
00:13:38,759 --> 00:13:40,139
DAW and both lead into the same
495
00:13:40,139 --> 00:13:40,620
house.
496
00:13:40,620 --> 00:13:43,659
The thing worth carrying away is
this the AI brain was never the
497
00:13:43,659 --> 00:13:45,340
hard part because everything
around it was.
498
00:13:45,340 --> 00:13:47,860
And that's exactly what the
platform takes off your plate.
499
00:13:47,860 --> 00:13:49,139
So here is my question for you,
500
00:13:49,139 --> 00:13:49,980
and I'd love to hear your
501
00:13:49,980 --> 00:13:50,659
answer.
502
00:13:50,659 --> 00:13:53,679
If you could hand one task to an
agent tomorrow, what would you
503
00:13:53,679 --> 00:13:56,659
hand over first and what would
you want it to leave alone?
504
00:13:56,659 --> 00:13:58,259
If that question got you
thinking?
505
00:13:58,259 --> 00:14:00,639
Leave your answer in the
comments and subscribe to the
506
00:14:00,639 --> 00:14:03,059
M365 FM podcast for more.
507
00:14:03,059 --> 00:14:04,740
Simply explained episodes that
508
00:14:04,740 --> 00:14:05,820
make this world a little less
509
00:14:05,820 --> 00:14:06,820
intimidating.
