WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=spNAUEgq_A8

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/spNAUEgq_A8.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:02.960 --> 00:00:14.360
Okay, so I'm going to be talking about domain-specific agents and why I really think that they are going to play an unbelievably important role in the future of AI

s2
00:00:14.360 --> 00:00:16.720
and in the future of how we build agents.

s3
00:00:16.720 --> 00:00:19.480
To get started real quick, my name's Justin Schrader.

s4
00:00:19.480 --> 00:00:23.480
Uh you can find me on X at JP Schrader.

s5
00:00:23.480 --> 00:00:29.920
And um I work at a small company called Standard Agents, which nobody's heard of right now cuz we're still kind of in stealth mode.

s6
00:00:29.920 --> 00:00:37.120
Um after this talk, if you're interested, feel free to reach out to me and uh I can let you know a little bit more.

s7
00:00:37.120 --> 00:00:41.360
Mostly, I'm known for doing a lot of different open-source projects.

s8
00:00:41.360 --> 00:00:45.800
Uh Dmux, which is a great multiplexer for all of your coding agents.

s9
00:00:45.800 --> 00:00:51.600
Uh ArrowJS, which is sort of like a UI framework, sort of like React for um the agentic era.

s10
00:00:51.600 --> 00:00:56.440
A bunch more that I won't get into, but you know, maybe check them out if you're interested.

s11
00:00:56.440 --> 00:00:57.400
Okay.

s12
00:00:57.400 --> 00:01:04.559
I think we can all agree that the moment in time that we are in is very similar to the Industrial Revolution.

s13
00:01:04.559 --> 00:01:07.880
Um in fact, it might be like an accelerated Industrial Revolution.

s14
00:01:07.880 --> 00:01:10.640
Maybe it's a bigger deal, but it's certainly not smaller.

s15
00:01:10.640 --> 00:01:14.480
I probably don't need to convince you of that if you're listening to one of these talks.

s16
00:01:14.480 --> 00:01:17.320
Um but that is the moment we find us find ourselves in.

s17
00:01:17.320 --> 00:01:24.800
So, I actually think it's helpful to go back and sort of look at what was the key catalyst of the Industrial Revolution.

s18
00:01:24.800 --> 00:01:29.960
And ultimately, it was that we learned how to harness energy with machines.

s19
00:01:29.960 --> 00:01:32.280
We learned how to harness energy with machines.

s20
00:01:32.280 --> 00:01:41.000
And what's interesting is that in this next era, we are essentially learning to harness intelligence with agents.

s21
00:01:41.000 --> 00:01:46.440
And agents, I think, can be thought of a little bit like the machine of yesteryear.

s22
00:01:46.440 --> 00:01:50.840
It's the thing that is going to use the intelligence.

s23
00:01:50.840 --> 00:01:55.000
Not so much us, but the agents.

s24
00:01:55.000 --> 00:02:04.800
What's interesting about that is I bet if I was in an actual room with you guys and and we all put up our hands, I bet a lot of you when I say what is an agent instantly

s25
00:02:04.800 --> 00:02:10.640
have examples that pop into mind, but also probably can't pull out a definition immediately.

s26
00:02:10.640 --> 00:02:23.000
Some of you maybe can, um but the reality is that we haven't even coalesced on a definition of what an agent is, even though we're well into the agentic era at this point.

s27
00:02:23.000 --> 00:02:25.160
And I think that's kind of interesting.

s28
00:02:25.160 --> 00:02:26.680
Um here's my definition.

s29
00:02:26.680 --> 00:02:39.680
You can feel free to agree with it or not, but agents are deterministic software that harness the non-deterministic results produced by models in pursuit of some desired objective.

s30
00:02:39.680 --> 00:02:55.280
Now, deterministic software might make you think more like a harness, and I actually think the distinction between an agent and a harness is really pedantic, not very helpful, um and for the most part, in most cases,

s31
00:02:55.280 --> 00:02:56.600
you can just conflate the two.

s32
00:02:56.600 --> 00:02:59.440
A harness is an agent and an agent is a harness, okay?

s33
00:02:59.440 --> 00:03:03.640
And for the for the purposes of this talk, we're going to go ahead and just move forward with that.

s34
00:03:03.640 --> 00:03:11.760
I think you could probably make some good arguments for why one is the other and vice versa, but really not important right now.

s35
00:03:11.760 --> 00:03:20.280
Now, if you did have some examples pop to mind, they might have been like Claude or Codex, um you know, OpenClaude, Hermes.

s36
00:03:20.280 --> 00:03:30.720
But you know what's interesting is I bet if you went out onto, you know, the the streets of corporate America in any city, maybe not San Francisco, but any city in America,

s37
00:03:30.720 --> 00:03:38.840
and you asked somebody just in an office building, could you name an agent by name?

s38
00:03:38.840 --> 00:03:42.880
I think some people are going to get Claude.

s39
00:03:42.880 --> 00:03:45.519
Some people might get Codex.

s40
00:03:45.519 --> 00:03:46.440
And that's about it.

s41
00:03:46.440 --> 00:03:50.239
I don't think hardly anybody's going to be getting OpenClaude or Hermes.

s42
00:03:50.239 --> 00:03:55.840
Uh Uh, and and really even Claude, I don't know that people would even know that that's an agent.

s43
00:03:55.840 --> 00:03:58.160
These things are not well understood.

s44
00:03:58.160 --> 00:04:03.720
And yet, what's so crazy is everybody is building agents.

s45
00:04:03.720 --> 00:04:07.360
I have a real estate agency down the street that's building agents.

s46
00:04:07.360 --> 00:04:13.040
I know in like independent private insurance brokers building their own agents.

s47
00:04:13.040 --> 00:04:17.720
I know Fortune 500 companies, lots of them, building their own custom agents.

s48
00:04:17.720 --> 00:04:22.000
Everybody is trying to build their own custom agents.

s49
00:04:22.000 --> 00:04:25.440
And I know people don't believe me, uh, but go talk to them.

s50
00:04:25.440 --> 00:04:27.000
Just go talk to people.

s51
00:04:27.000 --> 00:04:29.240
They are trying to build custom agents.

s52
00:04:29.240 --> 00:04:32.480
And I can't help but wonder why.

s53
00:04:32.480 --> 00:04:34.960
Nobody seems to be asking this question, why?

s54
00:04:34.960 --> 00:04:36.880
There's already AI everywhere.

s55
00:04:36.880 --> 00:04:43.840
You can get on ChatGPT all the way down to some open-source model from China on some, you know, rickety website.

s56
00:04:43.840 --> 00:04:47.480
There's everything in between, but still people want to build custom agents.

s57
00:04:47.480 --> 00:04:49.600
And ultimately, it comes down to integration.

s58
00:04:49.600 --> 00:04:54.919
Businesses want their data properly integrated into AI.

s59
00:04:54.919 --> 00:05:04.320
They They believe, and are probably right, that if they appropriately leverage AI, they're going to have these dramatic gains in their business and so on and so forth.

s60
00:05:04.320 --> 00:05:06.880
So, they need to figure out how to get integrated.

s61
00:05:06.880 --> 00:05:11.440
And building their own custom agents is obviously a way to do that.

s62
00:05:11.440 --> 00:05:16.400
And it's one of the first ways that they discover, um, as a mechanism for doing it.

s63
00:05:16.400 --> 00:05:20.560
The problem, though, is that agents are really hard.

s64
00:05:20.560 --> 00:05:27.960
You have to take very, very careful care of the agentic loop and make sure that it's properly orchestrated.

s65
00:05:27.960 --> 00:05:32.000
There are a ton of different provider abstractions you need to think about.

s66
00:05:32.000 --> 00:05:38.240
Um, fortunately, there's some good tools coming out around that, you know, like the the Vercel AI SDK is great.

s67
00:05:38.240 --> 00:05:43.280
Um, durable execution, you need to make sure if there's faults, we can pick back up.

s68
00:05:43.280 --> 00:05:48.800
These are relatively hard problems, um, especially if you're thinking about it at scale.

s69
00:05:48.800 --> 00:05:50.640
And the reality is there's just tons more.

s70
00:05:50.640 --> 00:05:54.560
There's all kinds of validations and stop conditions and so on and so forth.

s71
00:05:54.560 --> 00:06:01.600
And so what often happens is people do try to build their own custom agents, and they sort of work as a demo, but but not much more than that.

s72
00:06:01.600 --> 00:06:06.360
Um, and really it turns out that it's an absolute nightmare for people.

s73
00:06:06.360 --> 00:06:09.760
Um, building robust agents is just hard.

s74
00:06:09.760 --> 00:06:18.160
And if you go talk to anybody in an IT department, they are pulling their hair out because there are so many different concerns.

s75
00:06:18.160 --> 00:06:20.680
Um, there's no defined way to build an agent right now.

s76
00:06:20.680 --> 00:06:22.760
Like, actually no defined way.

s77
00:06:22.760 --> 00:06:29.320
The closest thing maybe is, uh, Eve that just came out from Vercel is maybe like the closest thing.

s78
00:06:29.320 --> 00:06:34.040
Um, but in reality, everybody's kind of coming up with their own way to do it.

s79
00:06:34.040 --> 00:06:38.800
Um, telemetry and observability on these agents is unbelievably hard, especially at scale.

s80
00:06:38.800 --> 00:06:50.560
Like, if you want to know exactly what is getting transmitted on every single step of every single turn of your agent, so that way you can diagnose it and fine-tune it and make sure things aren't going off the rails,

s81
00:06:50.560 --> 00:06:52.640
that is very hard to do.

s82
00:06:52.640 --> 00:06:54.480
Uh, agents are also not portable.

s83
00:06:54.480 --> 00:07:05.754
So, if I do get a good agent working, if I've managed to climb to the top of, you know, this mountain and I've got a good agent that's finally working well, well, it works well on my machine.

s84
00:07:05.754 --> 00:07:06.520
[laughter]

s85
00:07:06.520 --> 00:07:16.840
But if I try to pass it off to somebody else, there's a very high likelihood that between all of the environment variable configurations and and system requirements and and runtimes,

s86
00:07:16.840 --> 00:07:20.080
there's a good chance it's not going to run on that person's machine.

s87
00:07:20.080 --> 00:07:21.360
And they're not composable.

s88
00:07:21.360 --> 00:07:31.480
So, even if I get, you know, a really good chatbot working for my university, the chances that I'm going to then be able to reuse that for another thing

s89
00:07:31.480 --> 00:07:32.680
is very, very low.

s90
00:07:32.680 --> 00:07:35.280
I can't just easily share that.

s91
00:07:35.280 --> 00:07:41.720
So, what often happens is after a short pursuit towards agents, people kind of back away and say, "Okay, fine.

s92
00:07:41.720 --> 00:07:43.400
No more agents, no agents.

s93
00:07:43.400 --> 00:07:46.120
Instead, we're going to do the MCP thing.

s94
00:07:46.120 --> 00:07:48.000
We've heard about this, it works."

s95
00:07:48.000 --> 00:07:52.240
And sure enough, Model Context Protocol, it does work.

s96
00:07:52.240 --> 00:08:09.440
And really, it it works pretty well to take, you know, your corporate information like Zillow's information and then shove that into one of these really large pre-existing agents, something like Claude or ChatGPT, which I would consider a large general-purpose

s97
00:08:09.440 --> 00:08:10.320
agent.

s98
00:08:10.320 --> 00:08:12.280
Um and it and it sort of works like that.

s99
00:08:12.280 --> 00:08:14.400
Uh and and it works okay.

s100
00:08:14.400 --> 00:08:30.640
But if you take a look, this is actually from the MCP website, and if you take a look at what is supported in MCP clients around the world, you will notice that only one of these columns is actually filled out all the way down.

s101
00:08:30.640 --> 00:08:33.200
And that, of course, is tools.

s102
00:08:33.200 --> 00:08:40.080
So, MCP has become a de facto tool distribution mechanism for agents.

s103
00:08:40.080 --> 00:08:48.160
So, if I need to get my company's tools into that other agent, then MCP's a good way to do that.

s104
00:08:48.160 --> 00:08:54.680
It has not proven to be great at providing other value yet.

s105
00:08:55.000 --> 00:08:57.520
And frankly, tools are just not enough.

s106
00:08:57.520 --> 00:09:05.160
You know, I I I I like to joke that we didn't land a man on the moon by giving one guy a ton of tools.

s107
00:09:05.160 --> 00:09:10.880
That's not a realistic way to get a really large project done.

s108
00:09:10.880 --> 00:09:16.400
So, uh you know, maybe MCP's not the way, but aha, we have skills.

s109
00:09:16.400 --> 00:09:18.839
We have skills, and skills are great.

s110
00:09:18.839 --> 00:09:21.400
Um I I I actually do enjoy skills.

s111
00:09:21.400 --> 00:09:22.280
I'm sure you do, too.

s112
00:09:22.280 --> 00:09:24.560
We install them all the time for all kinds of things.

s113
00:09:24.560 --> 00:09:30.720
And fundamentally, what a skill is is a markdown file, which basically works as documentation.

s114
00:09:30.720 --> 00:09:38.840
Now, interestingly, there's lots of research out there that shows that if you use very many of these, it actually makes your agent substantially worse.

s115
00:09:38.840 --> 00:09:44.640
But, they do work as documentation for various complex things.

s116
00:09:44.640 --> 00:09:51.480
So, you know, back to the analogy of a man getting to the moon, it's a little bit like just giving this guy, you know, a ton of documentation.

s117
00:09:51.480 --> 00:09:56.600
And the documentation's going to help, but it's not the fundamental problem.

s118
00:09:56.600 --> 00:09:58.600
So, what's the fundamental problem?

s119
00:09:58.600 --> 00:09:59.880
Okay.

s120
00:09:59.880 --> 00:10:02.520
Let's build up a basic agent stack here.

s121
00:10:02.520 --> 00:10:03.480
Let's start with a model.

s122
00:10:03.480 --> 00:10:05.640
All agents start with a model.

s123
00:10:05.640 --> 00:10:07.160
Big one, small one, doesn't matter.

s124
00:10:07.160 --> 00:10:19.920
They Then, you have something like a system prompt on top of that, which tells the model what its role in the grand universe is, sort of like its its life objective.

s125
00:10:19.920 --> 00:10:24.040
Then, we have tools, the things that it can actually do, the effects it can take.

s126
00:10:24.040 --> 00:10:27.080
And then, skills would be layered on top of that.

s127
00:10:27.080 --> 00:10:30.360
And then, MCP would be layered on top of that.

s128
00:10:30.360 --> 00:10:34.040
And then, finally, you have all the messages from the conversation.

s129
00:10:34.040 --> 00:10:41.560
That is roughly the stack of information that gets passed along within the runtime of an agent.

s130
00:10:41.560 --> 00:10:48.400
And if you take a look here, almost all of it is context.

s131
00:10:48.560 --> 00:10:49.600
Basically, everything.

s132
00:10:49.600 --> 00:10:56.160
The system prompt, tools, skills, all of that is stuff that ends up in the context of the agent.

s133
00:10:56.160 --> 00:11:06.400
And so, basically, people are trying to solve the integration problem by working on the context or the model.

s134
00:11:06.400 --> 00:11:10.839
These are the two areas where we constantly see new advances.

s135
00:11:10.839 --> 00:11:17.720
We also see, you know, new new things come out like skills and and MCP, um new technologies, new protocols.

s136
00:11:17.720 --> 00:11:24.480
They are all coming out in the in the area of the context and the model.

s137
00:11:25.320 --> 00:11:28.080
So, how does it actually work then?

s138
00:11:28.080 --> 00:11:36.560
Well, basically, you work at a company, you occasionally need to do some business travel, so you've got a couple travel MCPs installed.

s139
00:11:36.560 --> 00:11:40.040
You've also got, you know, Figma and Playwright installed on yours.

s140
00:11:40.040 --> 00:11:43.120
And all of these are building up in that context layer.

s141
00:11:43.120 --> 00:11:53.160
And then you've got some, you know, Gmail MCPs to go check your mail for you and some Google Sheets to go fill out some other uh some other uh expense reports or something like that.

s142
00:11:53.160 --> 00:11:54.320
And then you've got skills.

s143
00:11:54.320 --> 00:11:58.200
You're a developer, so you've got some React fixers and linters.

s144
00:11:58.200 --> 00:12:03.320
This is actually, I think, like, the number one or the number two most popular MCP server that's out there.

s145
00:12:03.320 --> 00:12:10.080
Um maybe you've got uh Matt's uh grill me skill or or maybe you've got the GitHub skill.

s146
00:12:10.080 --> 00:12:14.880
And basically, what you're doing is you are inflating that context layer.

s147
00:12:14.880 --> 00:12:16.840
And we have a term for this in engineering.

s148
00:12:16.840 --> 00:12:18.440
It's called inheritance.

s149
00:12:18.440 --> 00:12:29.080
The idea of inheritance is you take an object and then you add more attributes to it to allow that one object to have other properties.

s150
00:12:29.080 --> 00:12:29.280
Right?

s151
00:12:29.280 --> 00:12:33.480
And that's exactly what we are doing here with an agent.

s152
00:12:33.480 --> 00:12:44.240
We're saying, "This agent is pretty good, but if we add all of these addish- additional extra layers, then the agent can do stuff that it previously couldn't do before."

s153
00:12:44.240 --> 00:12:46.560
That is exactly what inheritance is.

s154
00:12:46.560 --> 00:12:49.920
And the truth about inheritance is it works.

s155
00:12:49.920 --> 00:12:50.920
It does work.

s156
00:12:50.920 --> 00:12:55.200
That's why these things are out there and they are working.

s157
00:12:55.200 --> 00:13:01.120
But, there's an old saying, "Composition over inheritance."

s158
00:13:01.120 --> 00:13:04.320
And it turns out this this is as old as time.

s159
00:13:04.320 --> 00:13:07.040
Eventually, inheritance starts to break down.

s160
00:13:07.040 --> 00:13:13.960
Imagine, like, you know, okay, I've got five skills on uh ChatGPT or on or on Claude, excuse me.

s161
00:13:13.960 --> 00:13:15.920
And uh that works pretty well.

s162
00:13:15.920 --> 00:13:18.240
Now, what if I have 100 skills?

s163
00:13:18.240 --> 00:13:20.000
What if I have 1,000 skills?

s164
00:13:20.000 --> 00:13:25.360
There's some point at which I get diminishing returns from adding additional context.

s165
00:13:25.360 --> 00:13:27.280
That's That's just obvious.

s166
00:13:27.280 --> 00:13:30.040
We all kind of understand that implicitly.

s167
00:13:30.040 --> 00:13:32.080
So, is there an alternative?

s168
00:13:32.080 --> 00:13:35.000
Well, composition is the alternative to inheritance.

s169
00:13:35.000 --> 00:13:35.880
It looks something like this.

s170
00:13:35.880 --> 00:13:46.000
So, like imagine we have another little agent, and again, we're trying to provide Figma as an as a as a thing that can be done by our primary agent.

s171
00:13:46.000 --> 00:13:56.000
Well, what we could do is have a tiny little agent where the actual system prompt of the agent is written specifically to be a Figma agent.

s172
00:13:56.000 --> 00:13:57.680
It knows everything about Figma.

s173
00:13:57.680 --> 00:14:06.760
It it knows all of its all of its context, all of its API, all of the right places to click and the things to do and mouse movements to make and everything like that.

s174
00:14:06.760 --> 00:14:13.040
And then it has these precise tools that it needs to perform all of those actions, and nothing more.

s175
00:14:13.040 --> 00:14:14.480
Just that.

s176
00:14:14.480 --> 00:14:20.880
And then a very small message history, which just has to do with the Figma portion of this.

s177
00:14:20.880 --> 00:14:22.320
And then you can have more of these.

s178
00:14:22.320 --> 00:14:31.560
You can still have your Gmail and your travel and your Google Sheets and all that kind of stuff, but each of them is a separate isolated agent, a full agent.

s179
00:14:31.560 --> 00:14:33.600
Not just a little server with tools on it.

s180
00:14:33.600 --> 00:14:38.839
It's a full agent with its own message history, its own agentic loop.

s181
00:14:38.839 --> 00:14:43.160
And then above these, you have a coordinator.

s182
00:14:43.160 --> 00:14:51.800
And the communication mechanism for all of these small agents speaking to the larger agent above it is just English.

s183
00:14:51.800 --> 00:14:55.000
They just talk to each other the way a human does.

s184
00:14:55.000 --> 00:15:03.160
So, if the primary agent is saying, "Oh, I should I should check my mail to see if there's anything about going on a trip."

s185
00:15:03.160 --> 00:15:08.800
Well, it knows to go ask Gmail for any new uh emails about a trip.

s186
00:15:08.800 --> 00:15:15.360
Those funnel their way back up, says, "Oh, yeah, actually there's a trip coming up to Los Angeles this weekend."

s187
00:15:15.360 --> 00:15:20.600
And then it can go to our travel agent and start to make bookings.

s188
00:15:20.600 --> 00:15:26.880
That's kind of a a rough idea of how something like this could work.

s189
00:15:27.000 --> 00:15:28.960
And the reality is it does work.

s190
00:15:28.960 --> 00:15:33.080
And we know it works because this is actually how we got to the moon.

s191
00:15:33.080 --> 00:15:35.600
There were teams of experts.

s192
00:15:35.600 --> 00:15:41.520
Teams of experts with faces that looked like that and faces that looked like that.

s193
00:15:41.520 --> 00:15:44.560
Each of them with different skills and capabilities.

s194
00:15:44.560 --> 00:15:46.320
And faces that looked like that.

s195
00:15:46.320 --> 00:15:49.080
This is the Apollo 11 launch day.

s196
00:15:49.080 --> 00:15:50.440
And look right here, there's an agent.

s197
00:15:50.440 --> 00:15:53.035
I just found an agent sitting right there.

s198
00:15:53.035 --> 00:15:53.800
[laughter]

s199
00:15:53.800 --> 00:15:56.560
That brain of his is that's his LLM.

s200
00:15:56.560 --> 00:15:59.160
And here's his tools right there on the dashboard.

s201
00:15:59.160 --> 00:16:00.040
Those are the tools.

s202
00:16:00.040 --> 00:16:02.440
Now, he didn't have all the tools.

s203
00:16:02.440 --> 00:16:03.840
He just had those tools.

s204
00:16:03.840 --> 00:16:06.400
And he was really really really good at them.

s205
00:16:06.400 --> 00:16:08.200
And then look at that mouth.

s206
00:16:08.200 --> 00:16:09.800
That's the messages.

s207
00:16:09.800 --> 00:16:11.400
Uh we are used to this.

s208
00:16:11.400 --> 00:16:12.840
We can understand this.

s209
00:16:12.840 --> 00:16:14.560
It implicitly works.

s210
00:16:14.560 --> 00:16:18.800
It's almost a form of biomimicry for the agentic world.

s211
00:16:18.800 --> 00:16:21.120
Um it works.

s212
00:16:21.120 --> 00:16:23.320
And I call them domain-specific agents.

s213
00:16:23.320 --> 00:16:26.720
Um I don't think I was the first person to utter the words domain-specific agents.

s214
00:16:26.720 --> 00:16:29.520
Certainly not the first person to have this idea.

s215
00:16:29.520 --> 00:16:32.360
Um but that is what I want to talk to you about.

s216
00:16:32.360 --> 00:16:37.480
Agents that are just targeted to very specific domain.

s217
00:16:37.480 --> 00:16:43.320
And we over here at Standard Agents have been building this ecosystem for quite some time.

s218
00:16:43.320 --> 00:16:47.360
So we've gotten to have a really good inside look at how they actually work.

s219
00:16:47.360 --> 00:16:50.720
And I'm not ready to come out here and announce a product or anything like that.

s220
00:16:50.720 --> 00:16:52.640
Um but I can give you a little bit of a peek.

s221
00:16:52.640 --> 00:16:56.440
First of all, they are far more token efficient.

s222
00:16:56.440 --> 00:16:57.760
Far more token efficient.

s223
00:16:57.760 --> 00:17:03.839
We regularly see over 80% token efficiency for any given task.

s224
00:17:03.839 --> 00:17:10.280
Now, it's a little more complicated because you have to define those tasks a little bit more ahead of time.

s225
00:17:10.280 --> 00:17:19.600
But if you can have an agent portability where I can take that Gmail agent, squeeze it up, and then send it to somebody else, we can create an ecosystem

s226
00:17:19.600 --> 00:17:28.640
where we don't have to create every one of these skills and capabilities, but within that domain, you're going to get dramatic efficiency.

s227
00:17:28.880 --> 00:17:39.960
Um and part of the reason is, if you think about the way that the context work works, I don't need to have the entire context of the conversation when I make a choice to do something.

s228
00:17:39.960 --> 00:17:47.680
Instead, my primary coordinator level can just ask the Gmail uh hey, get that last email from Debbie.

s229
00:17:47.680 --> 00:17:49.920
And that is the totality of the context.

s230
00:17:49.920 --> 00:17:55.600
It literally just has the system message, its tools, and that message that came in.

s231
00:17:55.600 --> 00:18:05.640
And so, it is then able to perform this very targeted, very specialized, tiny little thing without all of the surrounding context.

s232
00:18:06.000 --> 00:18:12.520
It's also far more practical with small language models.

s233
00:18:12.520 --> 00:18:25.760
If you look at the difference in two models like DeepSeek uh V4 Flash and uh Fable 5, the cost difference is mind-boggling.

s234
00:18:25.760 --> 00:18:33.720
It is 137 times cheaper than Fable per task.

s235
00:18:33.720 --> 00:18:37.360
137 and 37 times.

s236
00:18:37.360 --> 00:18:51.760
Now, granted, if DeepSeek V4 Flash fails over and over and over again to do the job, then not only is it going to be, you know, not that much cheaper, it's also going to be much more annoying to use it.

s237
00:18:51.760 --> 00:19:00.320
But that's why domain-specific agents are so great, because you don't need to have the V4 Flash do everything.

s238
00:19:00.320 --> 00:19:06.440
Instead, it only needs to do the tasks that have been specifically picked for it to do.

s239
00:19:06.440 --> 00:19:11.440
And with a very minimal context, it can execute those very faithfully.

s240
00:19:11.440 --> 00:19:22.200
So, you get these dramatic cost reductions not only with the token efficiency, but also because you can use much smaller language models and even non-language models.

s241
00:19:22.200 --> 00:19:25.000
You can use image generation and diffusion models.

s242
00:19:25.000 --> 00:19:29.188
You can use all kinds of other models for smaller tasks.

s243
00:19:29.188 --> 00:19:30.080
[snorts]

s244
00:19:30.080 --> 00:19:34.200
You can also enforce really strict limits on the capabilities.

s245
00:19:34.200 --> 00:19:36.080
And I think you know what I'm talking about.

s246
00:19:36.080 --> 00:19:37.520
I'm talking about this.

s247
00:19:37.520 --> 00:19:41.400
Uh we are all flying awfully close to the sun nowadays.

s248
00:19:41.400 --> 00:19:44.920
We're everybody's just bypassing permissions left and right.

s249
00:19:44.920 --> 00:19:52.720
And of course, you have to because a coding agent with a big model can do anything.

s250
00:19:52.720 --> 00:19:56.400
And so, we use it to do everything.

s251
00:19:56.400 --> 00:20:01.640
In a world that would be powered by smaller domain-specific agents, those agents can't do everything.

s252
00:20:01.640 --> 00:20:06.600
They can only do the things that are already explicitly approved for them to do.

s253
00:20:06.600 --> 00:20:14.040
It doesn't mean that you still can't have permissions and permission dialogues, but you are opting into a much more controlled ecosystem.

s254
00:20:14.040 --> 00:20:25.519
And I promise you, when you explain that to Doug in IT, he puts his heart at ease understanding the difference between those two.

s255
00:20:25.519 --> 00:20:26.080
[snorts]

s256
00:20:26.080 --> 00:20:29.400
And uh fourth, they these have excellent scaling characteristics.

s257
00:20:29.400 --> 00:20:35.280
Because each of these agents is its own small little execution environment, you can parallelize them.

s258
00:20:35.280 --> 00:20:40.440
You can put them on the cloud very easily without needing like a giant VPC up there.

s259
00:20:40.440 --> 00:20:47.200
You can run thousands of instances all at the same time um in in all kinds of regions of the world.

s260
00:20:47.200 --> 00:20:52.880
They don't actually need to be uh you know, geographically co-located or anything like that.

s261
00:20:52.880 --> 00:20:56.880
Um so, they have very, very good scaling characteristics.

s262
00:20:56.960 --> 00:21:00.320
Unfortunately, they don't exist.

s263
00:21:00.440 --> 00:21:02.199
That's the downside.

s264
00:21:02.199 --> 00:21:02.720
[laughter]

s265
00:21:02.720 --> 00:21:06.680
These domain-specific agents don't really exist.

s266
00:21:06.680 --> 00:21:09.000
Um not in a big public way.

s267
00:21:09.000 --> 00:21:12.920
Like I said, here at Standard Agents, we have them.

s268
00:21:12.920 --> 00:21:19.880
We are working with them on a daily basis, but they are not out there in public very much yet.

s269
00:21:19.880 --> 00:21:21.960
However, that's changing.

s270
00:21:21.960 --> 00:21:23.920
That is going to change very quickly.

s271
00:21:23.920 --> 00:21:35.080
We're about halfway through 2026, and and I'm here to make a public prediction that I think as we roll on from from this point to the end of 2026,

s272
00:21:35.080 --> 00:21:43.800
we are going to see a dramatic uptick in people talking about building domain-specific agents, frameworks around them.

s273
00:21:43.800 --> 00:21:46.760
All kinds of things are coming down the pipe.

s274
00:21:46.760 --> 00:21:49.240
And it's not going to be a small trickle.

s275
00:21:49.240 --> 00:21:56.640
It's it's going to accelerate rapidly, and this will become a one of the main players in the agentic ecosystem.

s276
00:21:56.640 --> 00:22:02.080
And 2027, I would say, is basically the year of multi-agent orchestration.

s277
00:22:02.080 --> 00:22:05.280
That's another word you'll start to hear a lot, I think.

s278
00:22:05.280 --> 00:22:09.080
So, that's my big bold public prediction.

s279
00:22:09.080 --> 00:22:14.360
I was really excited just a few days ago when Vercel released Eve.

s280
00:22:14.360 --> 00:22:22.880
This is the first time I actually saw the term that I had been blasting out into the void come back and hit me in my own face.

s281
00:22:22.880 --> 00:22:29.960
The framework for building agents build a company brain, personal assistant, or domain-specific agent.

s282
00:22:29.960 --> 00:22:30.720
So, there we go.

s283
00:22:30.720 --> 00:22:35.000
About halfway through the year, we're going to start picking up steam.

s284
00:22:35.000 --> 00:22:37.320
That's my prediction.

s285
00:22:37.480 --> 00:22:38.960
And there's a number of reasons.

s286
00:22:38.960 --> 00:22:45.120
One of them is something that most people believe right now is that the cost of intelligence is going down.

s287
00:22:45.120 --> 00:22:49.080
That trend reversed in 2026, actually.

s288
00:22:49.080 --> 00:22:52.360
We track this on on a website.

s289
00:22:52.360 --> 00:22:54.800
Tokens are not getting cheaper anymore.

s290
00:22:54.800 --> 00:22:58.760
They are actually going up even when adjusted for IQ.

s291
00:22:58.760 --> 00:23:08.040
They're up 29% when you adjust for IQ just this year, halfway through the year, we're already up 30% and that can be caused by lots of different things.

s292
00:23:08.040 --> 00:23:21.200
Of course, um we've got this memory crunch and and you know, probably the long-term trend over a 10-year cycle or something is that intelligence will go down, but that does not mean that we need to be paying 137

s293
00:23:21.200 --> 00:23:25.920
times the cost for something that can be done just as effectively.

s294
00:23:25.920 --> 00:23:29.640
The problem is it's harder to break those things apart.

s295
00:23:29.640 --> 00:23:41.000
Now, if you don't account for IQ, tokens are up 76% this year, almost 100% increase in tokens just this year.

s296
00:23:41.000 --> 00:23:43.680
Um and we're we're not even halfway through it.

s297
00:23:43.680 --> 00:23:48.000
So, we are really trending upwards on on token costs.

s298
00:23:48.000 --> 00:23:52.800
So, anything we can do, especially with large businesses, to bring that down is going to be really important.

s299
00:23:52.800 --> 00:23:58.600
Um the other the other use case to really consider is putting AI in front of customers.

s300
00:23:58.600 --> 00:24:05.320
You can't put Fable in front of a customer, um unless that customer has a massive lifetime value.

s301
00:24:05.320 --> 00:24:07.120
It's just too expensive.

s302
00:24:07.120 --> 00:24:13.200
So, you need to find a way to create great efficacy while being efficient.

s303
00:24:13.200 --> 00:24:15.920
And domain-specific agents are going to be the way to do that.

s304
00:24:15.920 --> 00:24:22.280
So, I'm going to leave you here in momentarily, um but before I do, let's just dream a little bit.

s305
00:24:22.280 --> 00:24:31.960
Let me dig in a little bit deeper to how an agent could be orchestrated and what an ideal agent would actually look like.

s306
00:24:31.960 --> 00:24:35.120
And then, I promise to leave you alone.

s307
00:24:35.120 --> 00:24:35.680
Here we go.

s308
00:24:35.680 --> 00:24:39.600
So, remember we got that model and we got the system prompt.

s309
00:24:39.600 --> 00:24:42.600
And then, at the tool layer, let's break that apart a little bit.

s310
00:24:42.600 --> 00:24:44.320
On one hand, we have these like functions.

s311
00:24:44.320 --> 00:24:50.800
This would be like an actual function that can get executed, like write a file to the file system.

s312
00:24:50.800 --> 00:24:53.400
Then, we have prompts.

s313
00:24:53.400 --> 00:25:05.880
Prompts are a lot like the system prompt, but they are smaller individual prompts that can get injected and and sub prompts that can you know, you can run a function

s314
00:25:05.880 --> 00:25:07.960
that actually calls an LLM.

s315
00:25:07.960 --> 00:25:18.560
So, let's say I have a main agent running, but I want to use Nano Banana just to generate an image when I'm using GLM you know, 5.2 as my primary.

s316
00:25:18.560 --> 00:25:20.840
Well, you can just have a tool that's a prompt.

s317
00:25:20.840 --> 00:25:23.640
That would be really cool if you could do that.

s318
00:25:23.640 --> 00:25:33.840
And then another type of tool could be another full-blown agent, like a complete other domain specific agent could just be one of the tools.

s319
00:25:33.840 --> 00:25:35.720
So, that's the tool layer.

s320
00:25:35.720 --> 00:25:37.520
And then you have hooks.

s321
00:25:37.520 --> 00:25:38.720
Uh what are hooks?

s322
00:25:38.720 --> 00:25:48.600
Well, in this ideal world, a hook might be something that can kind of harness or change or mutate or perform side effects.

s323
00:25:48.600 --> 00:25:49.920
So, let me give you an example.

s324
00:25:49.920 --> 00:25:54.360
LLMs have no idea what time it is at any given point in time.

s325
00:25:54.360 --> 00:26:04.760
Turns out a really great way to tell them what time it is is you inject an artificial message or an artificial tool call in the message history.

s326
00:26:04.760 --> 00:26:04.812
So,

s327
00:26:04.812 --> 00:26:04.960
[snorts]

s328
00:26:04.960 --> 00:26:07.920
it looks like somebody just said, "Hey, what time it is?"

s329
00:26:07.920 --> 00:26:13.160
and the other person replied, "Oh, it's 6:45 p.m. Pacific time."

s330
00:26:13.160 --> 00:26:15.040
Pretty simple.

s331
00:26:15.040 --> 00:26:17.320
Um you can do that with a hook.

s332
00:26:17.320 --> 00:26:20.480
Or you could fire off some side effect using a hook.

s333
00:26:20.480 --> 00:26:23.960
So, this is an important piece of an agent.

s334
00:26:23.960 --> 00:26:26.680
And then finally, there's there's these agent rules.

s335
00:26:26.680 --> 00:26:28.480
And agent rules are kind of complicated.

s336
00:26:28.480 --> 00:26:33.240
It's like how many times should one side have a turn?

s337
00:26:33.240 --> 00:26:39.200
Like can it go on for 10,000 turns uh or 10,000 steps before it's turn is up?

s338
00:26:39.200 --> 00:26:41.840
You know, there's there's all kinds of interesting little rules.

s339
00:26:41.840 --> 00:26:48.840
You know, when it calls uh a tool, you know, is it required to validate the whole thing or not, you know?

s340
00:26:48.840 --> 00:26:54.680
All kinds of of very specific tools uh or or rules that that belong to a specific agent.

s341
00:26:54.680 --> 00:26:58.880
And altogether, if we bundle all that up, we would call that an agent.

s342
00:26:58.880 --> 00:27:00.640
But, it's kind of missing a couple things.

s343
00:27:00.640 --> 00:27:05.480
One, every agent should really have a file system.

s344
00:27:05.480 --> 00:27:15.960
If you've ever done this with ChatGPT or Claude or Codex, if you just ask it, you know, not inside of a project or anything like, "Hey, can you make me

s345
00:27:15.960 --> 00:27:20.320
a PDF for my my son's birthday party?"

s346
00:27:20.320 --> 00:27:21.920
Well, it'll do it.

s347
00:27:21.920 --> 00:27:25.000
And it'll store it in its own little file system.

s348
00:27:25.000 --> 00:27:33.360
So, the big labs have already realized that in order to create an effective chat interface, not to mention a big agent, it needs some sort of file system.

s349
00:27:33.360 --> 00:27:37.160
So, every agent should have its own little sandbox file system.

s350
00:27:37.160 --> 00:27:42.520
And also, every agent should have a sandboxed code execution location.

s351
00:27:42.520 --> 00:27:51.360
So, it can write files, it can run those files, and it can do that safely without exfiltrating anything, without interacting with an OS at a higher level.

s352
00:27:51.360 --> 00:27:56.840
That needs to be baked in as a primitive to every single domain-specific agent.

s353
00:27:56.840 --> 00:28:00.000
Okay, so, let's say that that's our ideal agent.

s354
00:28:00.000 --> 00:28:02.640
And now, let's talk about that little agent tool there.

s355
00:28:02.640 --> 00:28:03.800
What is that?

s356
00:28:03.800 --> 00:28:08.080
Well, those can be sub-agents, recursive sub-agents even.

s357
00:28:08.080 --> 00:28:14.360
You could have an agent that calls a sub-agent that calls sub-agents that call sub-agents.

s358
00:28:14.360 --> 00:28:18.560
Um, and there could be one or there could be many of these at different levels.

s359
00:28:18.560 --> 00:28:25.040
So, for example, you could have this coordinator agent that's at the very top, and then you could have a Salesforce agent.

s360
00:28:25.040 --> 00:28:27.400
And that agent knows Salesforce inside and out.

s361
00:28:27.400 --> 00:28:34.160
It knows all of its APIs, it knows has all the credentials to communicate with your Salesforce instance in all the appropriate ways.

s362
00:28:34.160 --> 00:28:38.120
And then, it needs to communicate with a Google Workspace agent.

s363
00:28:38.120 --> 00:28:39.920
So, it can do all kinds of stuff in there.

s364
00:28:39.920 --> 00:28:41.160
It can run spreadsheets.

s365
00:28:41.160 --> 00:28:45.600
I can say, "Hey, what are all my tops uh my my top sales people this year?"

s366
00:28:45.600 --> 00:28:51.480
Boom, it can look in Salesforce, it can coordinate with the sub agent, create a sheet for you, send that back.

s367
00:28:51.480 --> 00:28:52.840
Perfect.

s368
00:28:52.840 --> 00:28:55.360
But, maybe then you need to generate some assets.

s369
00:28:55.360 --> 00:29:00.480
So, the Salesforce agent actually has another sub agent that it can talk to at any time it wants.

s370
00:29:00.480 --> 00:29:02.720
And it's amazing at generating assets.

s371
00:29:02.720 --> 00:29:14.280
Maybe that sub agent doesn't just have like, you know, Codex image generation, maybe it has Nano Banana, maybe it has an SVG generator, all kinds of stuff, so that way it is an amazing

s372
00:29:14.280 --> 00:29:19.040
asset generator and performs some of its own reflection and QA.

s373
00:29:19.040 --> 00:29:25.880
And then our primary agent might need a whole legal team agent just so it can check the work that's coming out of these other ones.

s374
00:29:25.880 --> 00:29:32.040
And maybe the legal team agent really needs a GDPR compliance agent just for those European customers.

s375
00:29:32.040 --> 00:29:38.000
You know, the main one doesn't have all, you know, you don't we don't want to have 45 megabytes of context just on GDPR.

s376
00:29:38.000 --> 00:29:41.120
So, we make that a separate sub agent.

s377
00:29:41.120 --> 00:29:47.600
Uh so, you know, may and then maybe the legal team also needs like an OSHA compliance agent, which is also very complicated.

s378
00:29:47.600 --> 00:29:49.880
And so, it has a separate one for that.

s379
00:29:49.880 --> 00:29:51.000
You kind of get the idea.

s380
00:29:51.000 --> 00:30:04.560
You can end up with all kinds of highly efficient, small little agents that are all working together, but maintaining small minimal context windows all the way through.

s381
00:30:04.560 --> 00:30:08.120
That's the idea behind domain-specific agents.

s382
00:30:08.120 --> 00:30:09.440
So, thank you very much.

s383
00:30:09.440 --> 00:30:11.160
I appreciate you listening to my talk.

s384
00:30:11.160 --> 00:30:14.680
Again, uh Standard Agents is where we're working, standardagents.ai.

s385
00:30:14.680 --> 00:30:18.400
You can actually sign up on there uh for an early access.

s386
00:30:18.400 --> 00:30:21.920
Um we are slowly starting to roll this out to a few people.

s387
00:30:21.920 --> 00:30:31.120
Um if your business is super ambitious and really wants to uh try out small domain-specific agents, then um you know, you can write me info@standardagents.

s388
00:30:31.120 --> 00:30:32.920
And of course, I'd appreciate a follow.

s389
00:30:32.920 --> 00:30:34.320
Thank you so much.

s390
00:30:34.320 --> 00:30:36.320
Bye.