WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=GqoNrUz8hEU

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/GqoNrUz8hEU.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.640 --> 00:00:14.200
Hello.

s3
00:00:14.200 --> 00:00:16.640
Very sleepy crowd in the computer use room.

s4
00:00:16.640 --> 00:00:18.280
Have we all given up at this point?

s5
00:00:18.280 --> 00:00:19.920
Like, what's going on?

s6
00:00:19.920 --> 00:00:22.120
Uh thank you for coming in to my talk.

s7
00:00:22.120 --> 00:00:23.280
My name is Paul Klein.

s8
00:00:23.280 --> 00:00:28.440
I'm the founder of Browserbase, and I'm going to talk about bringing agents onto the World Wide Web.

s9
00:00:28.440 --> 00:00:36.960
If you're in this audience in this track, you've done computer use, you tried operator when it came out, and you're probably like, "Why isn't this happening yet?

s10
00:00:36.960 --> 00:00:38.080
This this seems obvious."

s11
00:00:38.080 --> 00:00:50.600
Well, we'll address some of the high-level needs of computer use to really serve what I think is the largest category of AI agents, the agents that actually go out and do work on your behalf in the real world.

s12
00:00:50.600 --> 00:01:00.840
We'll talk through some of the technical stuff, but really trying to focus on there's a huge model capabilities overhang in this category specifically, and you all here can hopefully solve it.

s13
00:01:00.840 --> 00:01:03.560
So, thank you for coming.

s14
00:01:03.600 --> 00:01:06.800
Well, of course, as we all know, the web wasn't built for agents.

s15
00:01:06.800 --> 00:01:13.400
It was built for people, and that becomes very challenging as we're building systems to try and interact with it and automate it.

s16
00:01:13.400 --> 00:01:21.120
So, when we're thinking about building agents that interact with people for systems, we have to really wonder, "Why wasn't it, you know, why was it built for us?

s17
00:01:21.120 --> 00:01:23.040
And and why is it struggle?"

s18
00:01:23.040 --> 00:01:27.080
When you've done any sort of automation in the past, you've run into so many roadblocks.

s19
00:01:27.080 --> 00:01:31.200
You know, the pages change, the web was built in a very context inefficient way.

s20
00:01:31.200 --> 00:01:33.360
It's a lot of text, a lot of tokens.

s21
00:01:33.360 --> 00:01:38.000
And when you're running any sort of browser agent or web agent right now, you get a broken browser that doesn't spin up.

s22
00:01:38.000 --> 00:01:39.760
You know, you have pages that don't work.

s23
00:01:39.760 --> 00:01:43.240
You have blockers or other sort of problems that really limit you.

s24
00:01:43.240 --> 00:01:47.920
And I actually started my career doing web automation, maintaining these scripts every single day.

s25
00:01:47.920 --> 00:01:48.880
It was very painful.

s26
00:01:48.880 --> 00:01:57.720
So, in my world, agents have made a huge advancement and allowed me to write durable web automation scripts, but we still haven't gone to agents yet.

s27
00:01:57.720 --> 00:01:59.600
And the question I want to ask is why?

s28
00:01:59.600 --> 00:02:05.040
We we're sitting here in this room we're thinking computer use we saw a year ago has progress stalled?

s29
00:02:05.040 --> 00:02:09.800
Like why why are web agents and browser extensions not as big as it could be?

s30
00:02:09.800 --> 00:02:12.959
And I think it's really a comes down to a few things.

s31
00:02:12.959 --> 00:02:16.600
You know, until recently the the bottom neck was the models.

s32
00:02:16.600 --> 00:02:21.920
The models one year ago really weren't good at long context horizon tasks.

s33
00:02:21.920 --> 00:02:25.560
But that's clearly been you know, solved in a major way.

s34
00:02:25.560 --> 00:02:31.600
Models can do more and more complex tasks than ever and of course in AI you always have to update your priors.

s35
00:02:31.600 --> 00:02:39.720
It's clear to me that anything I believe 6 months ago I have to revisit every single week because these models are progressing at an insanely fast pace.

s36
00:02:39.720 --> 00:02:41.480
So I don't think it's the models.

s37
00:02:41.480 --> 00:02:45.200
And especially models are now much better at using interfaces.

s38
00:02:45.200 --> 00:02:49.800
You know, we've seen this kind of capabilities improvement in computer use models.

s39
00:02:49.800 --> 00:02:54.320
In the last year a lot of investment was made in RL environments for coding.

s40
00:02:54.320 --> 00:03:00.320
And in the last 6 months months just as much investment has been made in RL environments for computer use.

s41
00:03:00.320 --> 00:03:03.680
And computer use models are getting better and you can see this in the evals.

s42
00:03:03.680 --> 00:03:10.760
When you train things on human trajectories in RL environments that model our real world, the real web, you can make better models.

s43
00:03:10.760 --> 00:03:13.360
So the models are getting there, I promise.

s44
00:03:13.360 --> 00:03:15.480
But the kind of so okay, so the models are good.

s45
00:03:15.480 --> 00:03:18.840
Why why do agents still struggle to use the web?

s46
00:03:18.840 --> 00:03:20.440
What are the what are the problems here?

s47
00:03:20.440 --> 00:03:24.959
If if it's a model problem, Docus says if the models were good enough diffusion would just happen.

s48
00:03:24.959 --> 00:03:27.880
Uh there's still a lot of work to be done here.

s49
00:03:27.880 --> 00:03:30.440
And and to me it's it's no longer just the models.

s50
00:03:30.440 --> 00:03:33.760
I'd argue that agents are missing the right harness and tools.

s51
00:03:33.760 --> 00:03:39.519
And if you aren't familiar with an agent harness, if you haven't been on Twitter in the last few weeks, it's the it's the scaffolding and systems around your model

s52
00:03:39.519 --> 00:03:41.840
that enable it to actually interact with the world.

s53
00:03:41.840 --> 00:03:52.560
Uh a lot of talks and talk about harness engineering we're going to spend a lot of time on today, but I really think you can invest a lot in a harness and get a lot more out of the models and extract that overhang out of the models.

s54
00:03:52.560 --> 00:04:02.160
You know, Karpathy actually tweeted this back in November 2023 and I thought it was just so forward-looking that this what he described, these systems around an LLM, it's the harness.

s55
00:04:02.160 --> 00:04:04.640
It's It's the the tools that it can access.

s56
00:04:04.640 --> 00:04:09.560
And if you fast-forward 3 years later, a lot of what we're doing every single day is building towards this.

s57
00:04:09.560 --> 00:04:15.120
A code interpreter for the LLM, audio and video input like screenshots, a browser, and other LLMs as subagents.

s58
00:04:15.120 --> 00:04:16.959
All of these principles have held true.

s59
00:04:16.959 --> 00:04:20.480
So, if you're ever wondering what should I build next, just go look at Karpathy's old talks.

s60
00:04:20.480 --> 00:04:23.200
He's a pretty good predictor of the future.

s61
00:04:23.200 --> 00:04:30.240
And you know, applying this to coding, we know that harnesses work really, really well with coding.

s62
00:04:30.240 --> 00:04:38.200
Uh on the the graphic on the right, you can see Factory when it compared to Claude Code using the same model but using their kind of custom uh harness.

s63
00:04:38.200 --> 00:04:47.480
And it turns out when you build a harness optimized for the domain that your agent is operating in, it can actually achieve, you know, above model results in that domain.

s64
00:04:47.480 --> 00:04:49.520
Harness engineering is a real thing.

s65
00:04:49.520 --> 00:04:51.760
And I'd say Cursor actually started this.

s66
00:04:51.760 --> 00:04:58.120
Cursor was the first one that was doing a model engineering or harness engineering on top of the original LLMs.

s67
00:04:58.120 --> 00:05:03.000
And a lot of what we've done at Browserbase with browser models has been, you know, harness engineering.

s68
00:05:03.000 --> 00:05:06.360
But and I think that like building a good harness is an engineering problem.

s69
00:05:06.360 --> 00:05:09.000
You don't have to be a lab to build a good harness.

s70
00:05:09.000 --> 00:05:11.400
And And a lot of us in the room maybe aren't working at labs.

s71
00:05:11.400 --> 00:05:15.640
Your company can make a great harness for your domain and actually improve model results.

s72
00:05:15.640 --> 00:05:16.919
You don't just have to wait for the models to catch up.

s73
00:05:16.919 --> 00:05:20.480
And once again, I believe the models are quite capable now.

s74
00:05:20.480 --> 00:05:25.480
And And if you look at this, you can see that it's not just Cursor, it's not just Factory.

s75
00:05:25.480 --> 00:05:32.960
You know, many, many different types of companies are building coding harnesses on top of models and over performing on the models' capabilities.

s76
00:05:32.960 --> 00:05:38.480
Now, it's not clear yet if custom harnesses are going to beat out durable, you know, RL'd models.

s77
00:05:38.480 --> 00:05:39.840
Uh but we're not going to debate that today.

s78
00:05:39.840 --> 00:05:45.440
We know that adding a harness on a model improves results, whether or not, you know, Claude Code will be the best harness ever or not.

s79
00:05:45.440 --> 00:05:46.720
I think that's a different conversation.

s80
00:05:46.720 --> 00:05:52.160
But you should still have some sort of harness on your model and measure the performance versus baseline model.

s81
00:05:52.160 --> 00:05:56.880
And what I really want to get back to is that there is a massive capabilities overhang in computer use.

s82
00:05:56.880 --> 00:06:00.560
The models are good enough, but we haven't done the engineering work to solve it.

s83
00:06:00.560 --> 00:06:06.080
And I love this great Brokman tweet where he says, "Whenever I don't use Codex for a task, I I ask myself why?"

s84
00:06:06.080 --> 00:06:09.120
And it feels like the task is an outside of the capabilities of the model.

s85
00:06:09.120 --> 00:06:10.400
The overhang is there.

s86
00:06:10.400 --> 00:06:13.160
The actual work we can do is missing.

s87
00:06:13.160 --> 00:06:24.640
And when you look at the amount of like task completion you can get with coding, it's so much higher than COA because we haven't actually really pushed the models far enough and given it the right tools.

s88
00:06:24.640 --> 00:06:33.040
So, to me not only is this important because I think that non-coding is a much bigger opportunity than it is coding.

s89
00:06:33.040 --> 00:06:39.520
If you look at this in recent slide, there's so many use cases that are in the non-coding domain that can benefit from computer use.

s90
00:06:39.520 --> 00:06:44.720
It's a problem worth investing in, and the wrong answer is to sit around and just wait for the models to get better.

s91
00:06:44.720 --> 00:06:47.240
You can actually solve this today.

s92
00:06:47.240 --> 00:06:57.320
Solving overhang is an engineering problem, and this is the work that we can do within our companies and within our agents, especially within the computer use domain, to build reliable web agents.

s93
00:06:57.560 --> 00:07:03.200
And when I think about browser agents that work, it it really comes down to three different types of things.

s94
00:07:03.200 --> 00:07:07.480
They're multimodal, they're harness engineered, and they have reliable infrastructure.

s95
00:07:07.480 --> 00:07:10.000
And I'll go through each of these three.

s96
00:07:10.000 --> 00:07:11.680
First, they're multimodal.

s97
00:07:11.680 --> 00:07:16.680
You no longer have to use a single model to actually interact with the task.

s98
00:07:16.680 --> 00:07:18.760
And we see this with coding agents all the time.

s99
00:07:18.760 --> 00:07:21.400
Sometimes you'll use a smarter model for a more complex page.

s100
00:07:21.400 --> 00:07:24.000
Sometimes a a dumber model for a simpler page.

s101
00:07:24.000 --> 00:07:28.800
And maybe you're using a combination of coding and computer use to actually power your agent.

s102
00:07:28.800 --> 00:07:30.520
This is a really important insight.

s103
00:07:30.520 --> 00:07:34.240
It turns out automating the web isn't always just clicking the button on the screen.

s104
00:07:34.240 --> 00:07:41.919
It might be intercepting the network requests and writing a coding agent or having coding write a script to actually replay those network requests.

s105
00:07:41.919 --> 00:07:49.920
The most reliable browser agents that we see in production right now are often writing code alongside using the browser to actually automate a task.

s106
00:07:49.920 --> 00:07:58.200
If you've done any sort of personal automation work in your life, you might see Cloud code output a script more often than using, you know, Cloud in Chrome because that's a very context-efficient

s107
00:07:58.200 --> 00:08:01.600
way to automate a repeatable task.

s108
00:08:01.600 --> 00:08:03.600
There's also harness engineering.

s109
00:08:03.600 --> 00:08:11.760
It turns out that sure we can, you know, write scripts and use models, but doing these things repeatedly, you want to benefit from things like memory and skills.

s110
00:08:11.760 --> 00:08:16.560
We launched something if we so-called browser.sh, which actually publishes skills for websites.

s111
00:08:16.560 --> 00:08:21.080
So, before your agent even goes to the website, it can observe what types of tasks it can do.

s112
00:08:21.080 --> 00:08:22.600
WebMCP is very useful for this.

s113
00:08:22.600 --> 00:08:26.040
It's part of pulling in existing knowledge to optimize a website.

s114
00:08:26.040 --> 00:08:29.800
Your agent doesn't have to discover something in the first place if it's done it before.

s115
00:08:29.800 --> 00:08:33.360
It can use its memory and its skills to actually make it better.

s116
00:08:33.360 --> 00:08:36.479
And you should think about trying to build in skills to your agents.

s117
00:08:36.479 --> 00:08:44.160
If your agent is using CLIs to control websites like the Playwright CLI, you could actually give it skills and context to be more effective there.

s118
00:08:44.160 --> 00:08:47.200
And this results in much more optimized token usage.

s119
00:08:47.200 --> 00:08:52.720
As if you're throwing everything on the page to a model, you're going to get sub-par results and it's going to cost you a lot.

s120
00:08:52.720 --> 00:09:02.000
The right harness should not only present the right tools, but present an optimized amount of tokens that are compressed to get exactly the right repeatable result every single time.

s121
00:09:02.000 --> 00:09:13.320
And finally, the infrastructure here is extremely important because when you're running browser agents in production, you want an environment that's going to work everywhere, every time.

s122
00:09:13.320 --> 00:09:15.080
And I think a lot of work still needs to be done here.

s123
00:09:15.080 --> 00:09:20.360
This is a lot of what our company does because computer use environments are pretty complex to scale up.

s124
00:09:20.360 --> 00:09:26.520
You know, it's funny when Open Cloud came out, everyone started buying Mac minis, which to me feels like an infrastructure problem, right?

s125
00:09:26.520 --> 00:09:37.000
You're running Open Cloud on a Mac mini in your house because that's the best way to run Mac OS that you can SSH into and then end up like solving the CAPTCHAs cuz of your home IP address.

s126
00:09:37.000 --> 00:09:41.120
That is not something you can do when you're building thousands of agents for customers in production.

s127
00:09:41.120 --> 00:09:46.160
I've yet to see a SOC 2 compliant Mac mini setup at scale, but please tell me afterwards if you found one.

s128
00:09:46.160 --> 00:09:47.720
I'm very curious about it.

s129
00:09:47.720 --> 00:09:51.360
The infrastructure problem that needs to be solved here is also an engineering problem.

s130
00:09:51.360 --> 00:09:55.840
And most importantly, the consistency in this infrastructure is important.

s131
00:09:55.840 --> 00:10:03.360
When your agent is running across a website multiple times, you want it to see the same inputs and outputs, the same page layout, the same size.

s132
00:10:03.360 --> 00:10:10.400
If your infrastructure renders a page in a mobile layout one time and then like in a desktop layout the second time, it's going to have inconsistent results.

s133
00:10:10.400 --> 00:10:18.760
Consistency in the infrastructure is the nice base layer on top of your harness and on top of your models to actually get good results with this.

s134
00:10:19.160 --> 00:10:21.360
I also think we have to improve the web itself.

s135
00:10:21.360 --> 00:10:26.880
So, there's a whole other side of this problem that's very interesting, which is like, how are we going to make it so the web works well with agents?

s136
00:10:26.880 --> 00:10:32.600
And I think this is arguably the harder challenge because we're not just engineering on our own systems anymore.

s137
00:10:32.600 --> 00:10:38.480
We have to be evangelists to the web and to the broader world that, "Hey, you want agents to come to your website."

s138
00:10:38.480 --> 00:10:42.160
So, accessibility is the first thing I want to talk about.

s139
00:10:42.160 --> 00:10:43.839
There's been a lot of really cool stuff here.

s140
00:10:43.839 --> 00:10:50.760
Now, when you look at what best-in-class browser agents are doing, they're not just consuming the raw DOM and HTML of the page anymore.

s141
00:10:50.760 --> 00:10:55.800
They're looking at sub, you know, subsections of that like the accessibility tree, the area tags.

s142
00:10:55.800 --> 00:11:01.680
These are labeled components of a page that can help show your agent where it needs to click and why.

s143
00:11:01.680 --> 00:11:04.280
Chrome just added WebMCP, which I think is really, really cool.

s144
00:11:04.280 --> 00:11:12.240
Websites can now publish MCP servers within their page that your agent can take advantage of without pre-installing the actual MCP.

s145
00:11:12.240 --> 00:11:22.520
It can now issue tool calls to a website like submit the registration form in a way that's not only context context context-efficient, but is website approved and blessed.

s146
00:11:22.520 --> 00:11:32.839
More and more work can go into accessibility and we've seen things like LLMs.text, skills.md, agents.md all being published alongside our websites.

s147
00:11:32.839 --> 00:11:37.080
We need to see more of that to build the agent-first web.

s148
00:11:37.080 --> 00:11:45.000
I think authentication is actually an even bigger problem here, too, because once your agent can actually go to a website, how can it log in on your behalf?

s149
00:11:45.000 --> 00:11:46.560
There's been a lot of different paradigms here.

s150
00:11:46.560 --> 00:11:51.000
Maybe you're just giving your agent your password, but doing that securely can be very challenging.

s151
00:11:51.000 --> 00:11:57.560
Maybe you're creating a service account for your agent where it has some limited access and you constantly have to give it new permissions.

s152
00:11:57.560 --> 00:12:08.160
You know, authentication for agents is the next thing to be solved once you solve the harnessing capability problems, and doing that securely where you can have a human loop approve certain actions on a website

s153
00:12:08.160 --> 00:12:11.960
is going to be a major challenge for unlocking computers for the enterprise.

s154
00:12:11.960 --> 00:12:17.040
The biggest gate to building agents that actually can work in prod is going to be the systems it has access to.

s155
00:12:17.040 --> 00:12:21.120
And authentication is something that needs to be solved in our industry to make this possible.

s156
00:12:21.120 --> 00:12:22.640
I've seen a lot of really cool stuff come out.

s157
00:12:22.640 --> 00:12:30.600
WorkOS just launched OffMD, which is a new way for your agent that goes to a website to find how to sign up on that website and get its own accounts.

s158
00:12:30.600 --> 00:12:38.480
And if you're building software now, you should think about what is my agent-first sign up and login flow look like because agents are going to be using your software whether you like it or not.

s159
00:12:38.480 --> 00:12:41.520
It's best to let them use it securely.

s160
00:12:41.760 --> 00:12:43.600
Finally, I want to talk about trust.

s161
00:12:43.600 --> 00:12:49.240
The web was built to stop bad bots, but now there's good agents and bad bots.

s162
00:12:49.240 --> 00:12:51.560
How do we delineate between the two?

s163
00:12:51.560 --> 00:12:59.839
And the CAPTCHA has been the tool in our tool chest for a very long time, but as we all know, CAPTCHAs are not as effective as we think against agents.

s164
00:12:59.839 --> 00:13:02.760
And trying to identify these good agents is very important.

s165
00:13:02.760 --> 00:13:14.400
There's been a lot of cool frameworks and work done on things like web bot off and in more authenticated ways to say this is my agent, it's coming from me, and you can follow me along on the web, but I still don't think we've solved the issue yet.

s166
00:13:14.400 --> 00:13:21.080
And a big unlock to agents accessing the web alongside authentication is actually how can we trust these agents?

s167
00:13:21.080 --> 00:13:30.520
And I think there needs to be almost like a Verisign moment for web agents where who can be the certificate issuer in saying my agent is trusted in this agent vendor is trusted?

s168
00:13:30.520 --> 00:13:32.040
Nobody's come out and done that yet.

s169
00:13:32.040 --> 00:13:35.480
I think those are really really big opportunities.

s170
00:13:35.680 --> 00:13:40.640
So, building reliable browser agents is is not a model problem.

s171
00:13:40.640 --> 00:13:48.120
It's an engineering problem that all of us can solve, but doing that engineering is is a full-time job.

s172
00:13:48.120 --> 00:13:51.440
And if you are working in the space, I'd love to meet you.

s173
00:13:51.440 --> 00:13:56.080
But if you aren't and you just want to build something that works, I might feel ideas.

s174
00:13:56.080 --> 00:13:58.760
You really don't have to reinvent the wheel here.

s175
00:13:58.760 --> 00:14:06.800
There's been a lot of stuff happening and it's a consortium of companies that are continuing to push the world forward on what's possible when you want to automate the web.

s176
00:14:06.800 --> 00:14:10.480
And I think there's like a few things here that are really important for a great solution, right?

s177
00:14:10.480 --> 00:14:13.320
It has to be a scalable platform that serves your infrastructure needs.

s178
00:14:13.320 --> 00:14:16.480
You know, you can want to run run one agent, but also thousands of agents.

s179
00:14:16.480 --> 00:14:20.120
And And the challenge is that those different levels of scale is very very important.

s180
00:14:20.120 --> 00:14:22.800
You want browser agents that are model agnostic.

s181
00:14:22.800 --> 00:14:26.600
As a developer, I don't want to be locked into a single model provider.

s182
00:14:26.600 --> 00:14:30.720
As models continually change and get better, I want to be able to move my agent around.

s183
00:14:30.720 --> 00:14:33.200
That's why you need model agnostic infrastructure.

s184
00:14:33.200 --> 00:14:35.560
You need somebody to solve agent identity.

s185
00:14:35.560 --> 00:14:47.240
Somebody who's going to go out and negotiate with the, you know, anti-bot providers of the world and say, "We are the platform for trusted agents and we are the ones that can help broker the access for your agents as you use the web."

s186
00:14:47.240 --> 00:14:49.360
And finally, you need observability.

s187
00:14:49.360 --> 00:14:57.120
When you're building these agents that go to any website in the world, you need to see where they're going and why and how that you can make sure that it's improving every iteration.

s188
00:14:57.120 --> 00:14:59.560
Every agent you run should get better every single time.

s189
00:14:59.560 --> 00:15:02.360
You need screen recordings, logs, network activity.

s190
00:15:02.360 --> 00:15:05.640
And you need to feed that back into your agent so it can self-improve.

s191
00:15:05.640 --> 00:15:08.680
We published something called Auto Browse earlier this year.

s192
00:15:08.680 --> 00:15:13.280
That's a really interesting way to see how is my agent able to improve itself over multiple loops.

s193
00:15:13.280 --> 00:15:20.000
And the the feed-in of data to that from observability is extremely important to make your agents get better over time.

s194
00:15:20.000 --> 00:15:22.520
And And that's, you know, what we're building here at BrowserBase.

s195
00:15:22.520 --> 00:15:30.520
You know, we power browser agents, web data extraction, and really all these use cases across the entire web to make your agents work well.

s196
00:15:30.520 --> 00:15:35.880
And what I've been extremely surprised by in building this company is the plethora of use cases.

s197
00:15:35.880 --> 00:15:47.040
Of course, there are the, you know, large AI native companies that use companies like BrowserBase to power the browser agents, but there's also all these little companies across the world that can benefit from automation.

s198
00:15:47.040 --> 00:15:54.360
And my core belief with this company is that solving computer use accelerates the diffusion of AI to the real economy.

s199
00:15:54.360 --> 00:16:03.560
And as much as I love our bubble here in San Francisco, the real economy is companies like the logistics company in Singapore, the bank in South Africa, or the lumber factory in Mexico.

s200
00:16:03.560 --> 00:16:09.760
These people are built on PHP websites with forums and human beings clicking buttons every single day.

s201
00:16:09.760 --> 00:16:18.440
That's a huge opportunity for you to go solve to build browser agents for them, and hopefully you can use the right infrastructure to power those things.

s202
00:16:18.440 --> 00:16:20.720
And and that's why we built BrowserBase agents, by the way.

s203
00:16:20.720 --> 00:16:27.839
This is our new product we launched yesterday, because we want to give everyone a battery-included agent and harness for everything they need to automate the web.

s204
00:16:27.839 --> 00:16:29.680
The goal here is that you shouldn't reinvent the wheel.

s205
00:16:29.680 --> 00:16:31.640
You shouldn't have to figure all this out and optimize it.

s206
00:16:31.640 --> 00:16:42.160
You should benefit from the platform scale that we've seen millions and millions of sessions every single month, and understand how we've solved the edge cases for you, so you don't have to solve them on your own.

s207
00:16:42.160 --> 00:16:43.320
I have a quick little demo here.

s208
00:16:43.320 --> 00:16:52.240
The way it works is instead of having to pull our tools together, you can actually put in a prompt, and we will stand up the harness, the runtime, the sandbox, the code execution,

s209
00:16:52.240 --> 00:16:56.800
the fetch, the search tools, and the models to actually accomplish a task for you.

s210
00:16:56.800 --> 00:17:06.360
And what's beautiful is as this agent is running, it's looking at its steps, and it's remembering what it can do, and learning from it, so it can do them again in the future.

s211
00:17:06.360 --> 00:17:09.319
The future for you is not having to reinvent the wheel every single time.

s212
00:17:09.319 --> 00:17:16.680
It's actually being able to use an agent that's purpose-built for browsing the web, and pull it in as a sub-agent of your larger agentic system.

s213
00:17:16.680 --> 00:17:19.480
This is not the main thing you should be focusing your time on.

s214
00:17:19.480 --> 00:17:25.720
You should be focusing your time on actually solving customer problems, not trying to rebuild the best in class browser agents.

s215
00:17:25.720 --> 00:17:27.480
The optimization feature is quite cool.

s216
00:17:27.480 --> 00:17:34.640
It's going to look back and actually understand, hey, how can I do this better after looking back at this this is this data feedback loop that I've talked about before.

s217
00:17:34.640 --> 00:17:38.320
And I think it's what makes agents really, really special.

s218
00:17:38.720 --> 00:17:41.440
I kind of want to end with this last note.

s219
00:17:41.440 --> 00:17:51.960
You know, based on the attendance in the room, I do think a lot of people have stepped back from computers because they've had so much challenges over the past year making browser agents work in production,

s220
00:17:51.960 --> 00:17:54.680
but I can tell you first hand from our customers we see it working.

s221
00:17:54.680 --> 00:18:01.880
And actually, I think 1 year from now this room is going to be overfilled with people because the models are getting better, the techniques are getting better, and the tools are getting better.

s222
00:18:01.880 --> 00:18:03.760
It's just on us to build better things.

s223
00:18:03.760 --> 00:18:04.640
Thank you all for having me today.

s224
00:18:04.640 --> 00:18:06.002
I really appreciate it.

s225
00:18:06.002 --> 00:18:08.002
[applause]

s226
00:18:23.687 --> 00:18:25.687
[music]