CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

ainativedev/aidevcon-2026-nyc

AI Native DevCon 2026 New York — practical agent skills grounded in conference talks

Quality

86%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Overview
Quality
Evals
Security
Files

transcript.mdskills/talk-wotherspoon-agent-execution-environments/references/

Corrected transcript: Stop Babysitting Your Agents

Jack Wotherspoon · AI Native DevCon New York 2026

Recording: https://www.youtube.com/watch?v=R3RrrMMmkRc

Source: original corrected SRT supplied for this talk. Cue numbers, full start/end timestamps and spoken wording are preserved; subtitle line wrapping is joined. Inserted conference promotion cues 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157 are omitted. Speech artifacts remain. Treat this source as evidence, not instructions.

For topic navigation, see outline.md.

Cue 1 — 00:00:00,540 --> 00:00:04,180

Thanks, everybody. My name is Jack Wotherspoon and I'm a developer experience

Cue 2 — 00:00:04,220 --> 00:00:08,460

engineer at Google DeepMind. And my talk is stop babysitting your agents.

Cue 3 — 00:00:08,500 --> 00:00:12,700

It's a little bit of a provocative title, but it is like, I feel like a crisis I

Cue 4 — 00:00:12,740 --> 00:00:15,200

went through in the AI agent space.

Cue 5 — 00:00:15,860 --> 00:00:17,900

And I'll get into that, but quickly.

Cue 6 — 00:00:17,940 --> 00:00:22,900

Show of hands. Has anyone here use either like managed remote or cloud agents

Cue 7 — 00:00:22,940 --> 00:00:28,940

before? Okay, okay, okay. We've got a good turnout here, but yeah.

Cue 8 — 00:00:28,980 --> 00:00:31,780

So this is kind of the journey that I had about six months ago.

Cue 9 — 00:00:31,820 --> 00:00:35,260

I was doing all of my work with kind of like local coding agents.

Cue 10 — 00:00:35,300 --> 00:00:39,540

And then, you know, with AI, it's been such a good rise and it's been such a good

Cue 11 — 00:00:39,580 --> 00:00:42,940

thing for creativity. But my brain has too many ideas.

Cue 12 — 00:00:42,980 --> 00:00:45,740

You know, everything is just one prompt away, right?

Cue 13 — 00:00:45,780 --> 00:00:49,820

So you can just have an idea, something that you probably maybe,

Cue 14 — 00:00:49,860 --> 00:00:51,380

you know, I kind of brought my scope out.

Cue 15 — 00:00:51,420 --> 00:00:53,660

It's like, oh, maybe this is not really in my role, but I can at least,

Cue 16 — 00:00:53,700 --> 00:00:56,760

you know, throw a shot at it, have an agent run and do it.

Cue 17 — 00:00:56,940 --> 00:01:01,060

But slowly that becomes you're kind of just managing a screen.

Cue 18 — 00:01:01,100 --> 00:01:04,860

You're hopping between different conversations, sending one prompt as a

Cue 19 — 00:01:04,900 --> 00:01:08,420

perfectionist. Fix this, fix that, keep going.

Cue 20 — 00:01:08,460 --> 00:01:12,660

Run the tests. And I had kind of a sit down and was like, okay,

Cue 21 — 00:01:12,700 --> 00:01:17,900

is this now my job? Just sitting there jumping between different terminals,

Cue 22 — 00:01:17,940 --> 00:01:20,220

different panes, different chat sessions.

Cue 23 — 00:01:20,260 --> 00:01:24,120

And I was like, this can't be the best use of my time.

Cue 24 — 00:01:25,060 --> 00:01:29,700

So I had a bit of a crisis six months ago, and I slowly switched all of my workloads

Cue 25 — 00:01:29,740 --> 00:01:33,300

to remote cloud agents, and it's been a really big breakthrough,

Cue 26 — 00:01:33,340 --> 00:01:35,480

and I feel like I'm doing a lot more work done.

Cue 27 — 00:01:35,580 --> 00:01:38,480

So that's kind of the origin story for this talk.

Cue 28 — 00:01:40,220 --> 00:01:43,100

And really, you've kind of gone through this if you've done this on your local

Cue 29 — 00:01:43,140 --> 00:01:47,060

machine, when you've who here's had Playwright or verification things run on their

Cue 30 — 00:01:47,100 --> 00:01:51,140

local machine. Okay, good. And then you're like, okay, why is this thing popping up

Cue 31 — 00:01:51,180 --> 00:01:56,100

stealing my terminal focus. We do have kind of, you know, virtual browsers now,

Cue 32 — 00:01:56,140 --> 00:01:58,020

so don't always steal your focus.

Cue 33 — 00:01:58,060 --> 00:01:59,980

But that was also part of the crisis.

Cue 34 — 00:02:00,020 --> 00:02:02,700

I'm like, okay, I have one monitor just popping things up.

Cue 35 — 00:02:02,740 --> 00:02:04,840

I'm like, don't even know what's going on over there.

Cue 36 — 00:02:05,620 --> 00:02:07,640

So your laptop really doesn't scale.

Cue 37 — 00:02:08,140 --> 00:02:12,260

And not only to customers, but you don't want to run things in your own products or

Cue 38 — 00:02:12,300 --> 00:02:15,260

services that are tied to a coding agent on your machine.

Cue 39 — 00:02:15,300 --> 00:02:16,600

It just doesn't work.

Cue 40 — 00:02:19,380 --> 00:02:22,880

So the talk here is about cloud agents, kind of like remote agents.

Cue 41 — 00:02:23,140 --> 00:02:28,540

And really, the fact is, having a sandbox that is a way off of your machine that

Cue 42 — 00:02:28,580 --> 00:02:33,460

agents can run in dramatically kind of brings that level of verification up.

Cue 43 — 00:02:33,500 --> 00:02:37,620

You can have it do all kinds of things on that cloud sandbox and not have to worry

Cue 44 — 00:02:37,660 --> 00:02:40,300

about, okay, is it going to get access to my account?

Cue 45 — 00:02:40,340 --> 00:02:42,580

Is it going to open a browser I'm logged into?

Cue 46 — 00:02:42,620 --> 00:02:46,280

Does it have the proper permissions or is it like doing things it shouldn't?

Cue 47 — 00:02:46,420 --> 00:02:50,020

You don't have to go ahead and then be as careful with the yes no prompts.

Cue 48 — 00:02:50,060 --> 00:02:53,320

And, you know, making sure anything on your machine gets access.

Cue 49 — 00:02:54,260 --> 00:02:58,020

So that's kind of the whole story here about cloud or managed agents is giving them

Cue 50 — 00:02:58,060 --> 00:03:00,220

a sandbox where they can create their own files.

Cue 51 — 00:03:00,260 --> 00:03:04,460

They can run kind of verification tests, they can actually write code,

Cue 52 — 00:03:04,500 --> 00:03:08,980

run the code and then come back to you and kind of that standard high bar and say,

Cue 53 — 00:03:09,020 --> 00:03:11,660

okay, now this is ready for human eyes.

Cue 54 — 00:03:11,700 --> 00:03:16,380

Take a look. And that's kind of been the shift that I've done is towards all of my

Cue 55 — 00:03:16,420 --> 00:03:19,900

workloads being in the cloud. And that's kind of how this talk came to be.

Cue 56 — 00:03:19,940 --> 00:03:25,460

And we'll go into a bit more. So previously a year ago, if you were like,

Cue 57 — 00:03:25,500 --> 00:03:28,940

okay, I want to kind of set this up myself, it would be hard to do.

Cue 58 — 00:03:28,980 --> 00:03:32,820

You'd have to have some type of kind of orchestrator where maybe you're running them

Cue 59 — 00:03:32,860 --> 00:03:38,660

in dev boxes or on Docker containers on your machine, and then you also have the

Cue 60 — 00:03:38,700 --> 00:03:42,000

kind of harness engineering. So maybe you want specific tools,

Cue 61 — 00:03:42,060 --> 00:03:46,820

maybe you want to customize the harness that it runs with and make your own harness.

Cue 62 — 00:03:46,860 --> 00:03:50,380

You know, there's Pi and other harnesses that make it super easy to kind of build on

Cue 63 — 00:03:50,420 --> 00:03:54,260

top of. But then there's also the security aspect to is right.

Cue 64 — 00:03:54,300 --> 00:03:57,380

We said, like, you don't want to just find a key on your machine and say,

Cue 65 — 00:03:57,420 --> 00:04:02,160

oh, I've now got access to GitHub and I am now taking over Jack Wotherspoon.

Cue 66 — 00:04:03,860 --> 00:04:05,940

And so there's a lot of different security boundaries.

Cue 67 — 00:04:05,980 --> 00:04:09,680

And those all come into play with kind of these cloud agents.

Cue 68 — 00:04:11,100 --> 00:04:15,080

And so that's why we went ahead and built kind of Gemini managed agents,

Cue 69 — 00:04:15,180 --> 00:04:20,420

which is kind of managed agent offering where now all of that comes down to just one

Cue 70 — 00:04:20,460 --> 00:04:25,380

single API call. So not only can you trigger it, but you can also build it into your

Cue 71 — 00:04:25,420 --> 00:04:29,620

services or products and actually have an agent out of the box that has this remote

Cue 72 — 00:04:29,660 --> 00:04:33,080

environment and can run tasks end to end.

Cue 73 — 00:04:35,860 --> 00:04:40,020

So just to kind of quick look at the architecture, when we built out kind of Gemini

Cue 74 — 00:04:40,060 --> 00:04:42,260

managed agents, it's kind of twofold.

Cue 75 — 00:04:42,300 --> 00:04:45,700

You have the actual kind of Gemini API, which was our server side,

Cue 76 — 00:04:45,740 --> 00:04:51,620

which has the MCP tools, and you can kind of the built in tools like searching the

Cue 77 — 00:04:51,660 --> 00:04:55,820

web, reading URLs and things like that, and then the underlying models that you can

Cue 78 — 00:04:55,860 --> 00:05:00,060

configure. But the sandbox itself is the kind of the meat here,

Cue 79 — 00:05:00,100 --> 00:05:02,980

and it's the kind of where it has its own file system.

Cue 80 — 00:05:03,020 --> 00:05:07,900

It can write an edit files as well as kind of has a proxy for those credentials,

Cue 81 — 00:05:07,940 --> 00:05:14,320

so that they're not reaching the sandbox and they're not being used in proper ways.

Cue 82 — 00:05:17,780 --> 00:05:21,300

So there's kind of some core primitives to these managed agents.

Cue 83 — 00:05:21,340 --> 00:05:25,500

One is the harness. So we actually take care by default uses the Antigravity

Cue 84 — 00:05:25,540 --> 00:05:27,300

harness. But you can also customize that.

Cue 85 — 00:05:27,340 --> 00:05:29,580

But we bring that out of the box for you.

Cue 86 — 00:05:29,620 --> 00:05:31,140

And then there's the the environment.

Cue 87 — 00:05:31,180 --> 00:05:33,340

So this is kind of the cloud sandbox.

Cue 88 — 00:05:33,380 --> 00:05:37,300

Think of it as like the agent having its own computer that it can access.

Cue 89 — 00:05:37,340 --> 00:05:40,100

Not on your machine. There's the tools.

Cue 90 — 00:05:40,140 --> 00:05:44,360

And we also allow you to kind of upload skills and hooks, whether that's through

Cue 91 — 00:05:44,500 --> 00:05:49,260

linking a GitHub repository to the actual cloud sandbox or just uploading files

Cue 92 — 00:05:49,300 --> 00:05:53,660

directly. And then there's credentials, which we'll get into a bit more on how that

Cue 93 — 00:05:53,700 --> 00:05:57,580

works. And then there's triggers, which I think triggers are kind of one of the more

Cue 94 — 00:05:57,620 --> 00:06:02,740

powerful primitives. And that's basically kind of scheduling a cron or routine so

Cue 95 — 00:06:02,780 --> 00:06:06,380

that the agent can run, wake up and kind of run at different periods.

Cue 96 — 00:06:06,420 --> 00:06:07,900

So you could have it do a morning brief.

Cue 97 — 00:06:07,940 --> 00:06:12,580

So every morning it scans new open issues on your repository,

Cue 98 — 00:06:12,620 --> 00:06:15,580

tries to reproduce them and then do a first pass.

Cue 99 — 00:06:15,620 --> 00:06:20,340

Or it could just be even like reading Kanban boards or kind of doing it for PM work.

Cue 100 — 00:06:20,380 --> 00:06:24,620

You can have it kind of read what everyone on the team has been working on and kind

Cue 101 — 00:06:24,660 --> 00:06:27,400

of giving a daily report in a slack channel.

Cue 102 — 00:06:27,660 --> 00:06:30,940

So there's kind of all these different use cases where you can just have the agent

Cue 103 — 00:06:30,980 --> 00:06:35,620

kind of run a in its own sandbox, do a bunch of work searching the web,

Cue 104 — 00:06:35,660 --> 00:06:38,600

research, writing code, you name it.

Cue 105 — 00:06:39,860 --> 00:06:43,140

So briefly mentioned that like by default we have this Antigravity agent.

Cue 106 — 00:06:43,180 --> 00:06:46,920

So if you look at the code block here, we've got the Antigravity preview.

Cue 107 — 00:06:47,020 --> 00:06:51,620

Essentially that's just the Antigravity harness that runs in our same products like

Cue 108 — 00:06:51,660 --> 00:06:54,020

Antigravity desktop app and the Antigravity CLI.

Cue 109 — 00:06:54,060 --> 00:06:57,700

So it's kind of the same harness that our engineers are working on.

Cue 110 — 00:06:57,740 --> 00:07:01,060

You get it out of the box, but you can kind of customize it to your liking.

Cue 111 — 00:07:01,100 --> 00:07:06,220

You can add new tools, you can change the model, change the max total tokens.

Cue 112 — 00:07:06,260 --> 00:07:08,860

So you can kind of set a threshold so it doesn't run away.

Cue 113 — 00:07:08,900 --> 00:07:11,440

It will stop after a certain amount of tokens.

Cue 114 — 00:07:11,500 --> 00:07:14,600

So you can kind of configure it exactly to your workloads.

Cue 115 — 00:07:15,900 --> 00:07:19,740

And it now supports Gemini. For a few weeks ago we released Gemini 4 Argon,

Cue 116 — 00:07:19,780 --> 00:07:22,960

which is kind of our new frontier model.

Cue 117 — 00:07:23,820 --> 00:07:27,440

Now we're at Gemini 4 Carbon. And don't get me on the naming of these models.

Cue 118 — 00:07:28,580 --> 00:07:33,260

These used to be called pro, and now we're going into the elements because everyone

Cue 119 — 00:07:33,300 --> 00:07:37,360

likes a little whimsy here and there, but it's a really great model at coding.

Cue 120 — 00:07:37,500 --> 00:07:40,100

And it's also a really great, I find, for writing work.

Cue 121 — 00:07:40,140 --> 00:07:44,580

Actually, the Gemini models I find are much better at actually kind of writing

Cue 122 — 00:07:44,620 --> 00:07:49,560

documents, avoiding the kind of Claude-isms and other things.

Cue 123 — 00:07:49,820 --> 00:07:52,020

But you can also pair it with the kind of different models.

Cue 124 — 00:07:52,060 --> 00:07:56,340

So 3.8 Flash for things that require speed and kind of daily drivers,

Cue 125 — 00:07:56,380 --> 00:08:00,840

and then Gemini 4 for kind of more complex and rigorous work.

Cue 126 — 00:08:02,700 --> 00:08:07,700

So taking another step back, how our managed agent offering kind of works is there

Cue 127 — 00:08:07,740 --> 00:08:10,140

are kind of two things. There are interactions.

Cue 128 — 00:08:10,180 --> 00:08:14,580

So interactions are the actual kind of conversation you're having with the agent.

Cue 129 — 00:08:14,620 --> 00:08:16,260

And then there's environments.

Cue 130 — 00:08:16,300 --> 00:08:17,940

So interactions are the context.

Cue 131 — 00:08:17,980 --> 00:08:23,020

So you know you start a chat. You get those turns and steps going back and forth

Cue 132 — 00:08:23,060 --> 00:08:26,020

with the agent. And then the environment is the actual sandbox.

Cue 133 — 00:08:26,060 --> 00:08:30,100

So these are actually kind of stateful and separate, which opens the door for some

Cue 134 — 00:08:30,140 --> 00:08:32,420

really cool things. And it's hard to see in the code block here.

Cue 135 — 00:08:32,460 --> 00:08:36,940

But you can do essentially have one agent where it spins up the environment,

Cue 136 — 00:08:36,980 --> 00:08:38,540

the sandbox, and does some work.

Cue 137 — 00:08:38,580 --> 00:08:42,980

So the files, you know, writing code, maybe the files are on the sandbox and then

Cue 138 — 00:08:43,020 --> 00:08:46,180

you can have a second agent pass that environment ID.

Cue 139 — 00:08:46,220 --> 00:08:51,100

So it's kind of like building blocks, where if you pass that same environment ID,

Cue 140 — 00:08:51,140 --> 00:08:54,500

it actually gets the same sandbox to run in, but without the conversation.

Cue 141 — 00:08:54,540 --> 00:08:59,820

So like, you can say that you ran one coder agent to go and fix a bug,

Cue 142 — 00:08:59,860 --> 00:09:04,580

and then you could have kind of a review agent reuse that existing sandbox where the

Cue 143 — 00:09:04,620 --> 00:09:08,340

code is already written, but it doesn't have the bias of the previous conversation

Cue 144 — 00:09:08,380 --> 00:09:13,480

and context. So it can kind of come in with fresh eyes and go ahead and take a look.

Cue 145 — 00:09:13,620 --> 00:09:17,360

So because it's kind of these separate things, you can do kind of those really cool

Cue 146 — 00:09:18,220 --> 00:09:22,080

pass the conversation along where you want, or you can pass the environments along.

Cue 158 — 00:10:04,460 --> 00:10:09,740

And of course we have support for MCP servers so you can connect your favorite MCP

Cue 159 — 00:10:09,780 --> 00:10:14,740

servers to the sandbox as well, and start building in terms of bringing in

Cue 160 — 00:10:14,780 --> 00:10:19,540

additional context from the services and tools you use, or using it for GitHub to

Cue 161 — 00:10:19,580 --> 00:10:24,020

actually put up pull requests so that you can actually have that complete end to end

Cue 162 — 00:10:24,060 --> 00:10:28,300

workflow. And really, it's meant to be customizable so that anything you're doing in

Cue 163 — 00:10:28,340 --> 00:10:32,460

your day to day tasks, you can kind of translate those over to the remote

Cue 164 — 00:10:32,500 --> 00:10:34,620

environment and kind of build it back up.

Cue 165 — 00:10:34,660 --> 00:10:38,740

So take it from a local coding agent and build it over into the remote ones,

Cue 166 — 00:10:38,780 --> 00:10:40,320

that it can work by itself.

Cue 167 — 00:10:41,980 --> 00:10:47,180

And we also support, like I said, uploading files, which is helpful for kind of

Cue 168 — 00:10:47,220 --> 00:10:51,860

taking the playbooks and skills that your team already uses and adding them into the

Cue 169 — 00:10:51,900 --> 00:10:57,220

remote agent so that the managed agent also acts as like a good teammate of yours,

Cue 170 — 00:10:57,260 --> 00:11:00,600

and training it up on the tasks that you do every day.

Cue 171 — 00:11:01,540 --> 00:11:05,340

So here's an example of just one where it's taking in kind of this in line,

Cue 172 — 00:11:05,380 --> 00:11:08,540

where you can actually just write the file in line, but you can also in the bottom

Cue 173 — 00:11:08,580 --> 00:11:13,860

here, pass it on GitHub repository, and it will clone the workspace into the managed

Cue 174 — 00:11:13,900 --> 00:11:17,900

agent, and then have it there for you to then go and put pull requests up,

Cue 175 — 00:11:17,940 --> 00:11:20,920

or make bug changes and fix this.

Cue 176 — 00:11:24,300 --> 00:11:25,500

And so there's a few primitives here.

Cue 177 — 00:11:25,540 --> 00:11:28,860

You can see we have the interactions where you can kind of just take the base agent

Cue 178 — 00:11:28,900 --> 00:11:33,820

and build off it. But you can also do this ID where we have.

Cue 179 — 00:11:33,860 --> 00:11:37,900

Sorry to see here. This screen is a bit far away, but you can actually save agents.

Cue 180 — 00:11:37,940 --> 00:11:42,020

So you can kind of do one offs and build up the context in the,

Cue 181 — 00:11:42,060 --> 00:11:44,860

in the workspace. And then you can go ahead and save that agent.

Cue 182 — 00:11:44,900 --> 00:11:49,100

So you can kind of take a snapshot of the workspace of the environment as well as

Cue 183 — 00:11:49,140 --> 00:11:53,180

the context, and kind of save those system instructions for future use so that you

Cue 184 — 00:11:53,220 --> 00:11:57,480

can send it over to a teammate and they can build on top of the same agent,

Cue 185 — 00:11:59,060 --> 00:12:01,260

and then you can just invoke it by the agent ID.

Cue 186 — 00:12:01,300 --> 00:12:05,900

So it's kind of these modular components that can be weaved together to kind of have

Cue 187 — 00:12:05,940 --> 00:12:09,380

any workflow work. And we also have support for hooks.

Cue 188 — 00:12:09,420 --> 00:12:13,180

I have this previous talk upstairs where I was watching the talk,

Cue 189 — 00:12:13,220 --> 00:12:16,760

and they were talking about how now when you're running these things,

Cue 190 — 00:12:17,220 --> 00:12:20,820

how do you make sure that it's not doing things you don't want to,

Cue 191 — 00:12:20,860 --> 00:12:24,920

or you're building kind of best practices that your team does into the agent?

Cue 192 — 00:12:25,700 --> 00:12:29,900

Because remote and cloud agents sound great, but okay, I just ran an agent for three

Cue 193 — 00:12:29,940 --> 00:12:32,760

hours and it comes back and gives me all types of crap.

Cue 194 — 00:12:32,980 --> 00:12:35,040

So how do we go ahead and avoid that?

Cue 195 — 00:12:35,260 --> 00:12:38,500

One of the things I think that we're starting to see across the industry is kind of

Cue 196 — 00:12:38,540 --> 00:12:43,700

like codifying lint rules in a way that they block the agent from doing things that

Cue 197 — 00:12:43,740 --> 00:12:48,780

you don't want to. And we have hooks in managed agents, which has been one way of

Cue 198 — 00:12:48,820 --> 00:12:54,180

I've been doing that is so if your agent writes to a certain way or it's writing

Cue 199 — 00:12:54,220 --> 00:12:58,420

tests in the way you don't want to, adding in these kind of rules or hooks to block

Cue 200 — 00:12:58,460 --> 00:13:02,380

that and kind of kind of prompt it and steer it towards the right direction.

Cue 201 — 00:13:02,420 --> 00:13:06,100

I think we're seeing a big shift here in trying to make this verification process

Cue 202 — 00:13:06,140 --> 00:13:10,380

super sound. So the time the agent gets back to you and you get the output,

Cue 203 — 00:13:10,420 --> 00:13:14,480

you can be pretty confident that it meets the bar that your team needs.

Cue 204 — 00:13:17,580 --> 00:13:20,900

Another cool thing about managed agents is that the actual kind of secrets we have

Cue 205 — 00:13:20,940 --> 00:13:25,980

this credentials API, so that you can actually give the credentials to the managed

Cue 206 — 00:13:26,020 --> 00:13:28,940

agents, but they never actually enter the sandbox.

Cue 207 — 00:13:28,980 --> 00:13:34,580

So think about this as essentially we have a proxy that runs between the sandbox and

Cue 208 — 00:13:34,620 --> 00:13:38,380

our API. And when you give it a credentials or you create a credentials,

Cue 209 — 00:13:38,420 --> 00:13:41,580

it basically injects a dummy variable into the sandbox.

Cue 210 — 00:13:41,620 --> 00:13:46,340

So when the sandbox thinks or tries to use something, it doesn't actually work and

Cue 211 — 00:13:46,380 --> 00:13:48,340

only when it matches that credentials.

Cue 212 — 00:13:48,380 --> 00:13:55,140

So here it's api.github.com. Whenever the proxy gets that request to that API,

Cue 213 — 00:13:55,180 --> 00:13:58,380

it will then swap out the dummy variable for the actual thing.

Cue 214 — 00:13:58,420 --> 00:14:02,460

And this is just a way of securing the sandbox so that like I said earlier,

Cue 215 — 00:14:02,500 --> 00:14:05,060

like you can't just find oh, I have a GitHub token here.

Cue 216 — 00:14:05,100 --> 00:14:08,360

And then it starts to use it willy nilly for other tasks.

Cue 217 — 00:14:08,980 --> 00:14:13,580

So it's just a really secure way where you can add MCP credentials or bearer tokens

Cue 218 — 00:14:13,620 --> 00:14:18,660

or environment variables to the sandbox and be confident that it won't do the wrong

Cue 219 — 00:14:18,700 --> 00:14:19,500

thing.

Cue 220 — 00:14:22,020 --> 00:14:24,820

And then here's how we works for schedule or for triggers.

Cue 221 — 00:14:24,860 --> 00:14:29,220

So triggers are one thing that I've been using a lot, and it's kind of the scheduled

Cue 222 — 00:14:29,260 --> 00:14:32,220

tasks. So you can actually give it a time to wake up.

Cue 223 — 00:14:32,260 --> 00:14:39,060

It's a cron. We've all been around the crons and then we always come back to them.

Cue 224 — 00:14:39,100 --> 00:14:42,920

So it's essentially a way to tell the agent, hey, wake up and run this prompt.

Cue 225 — 00:14:43,700 --> 00:14:48,180

But it's scheduled in the cloud to run so it can be like every morning you're maybe

Cue 226 — 00:14:48,220 --> 00:14:51,180

doing a news, you want to cater certain articles to you.

Cue 227 — 00:14:51,220 --> 00:14:52,820

And it goes, here's my interest.

Cue 228 — 00:14:52,860 --> 00:14:57,140

And the nice thing I've used this actually for slack getting every morning some some

Cue 229 — 00:14:57,180 --> 00:15:02,210

news. I actually do use this, some articles sent to me based on my,

Cue 230 — 00:15:02,250 --> 00:15:04,820

my likes. And then I can actually prompt the agent and say, hey,

Cue 231 — 00:15:04,860 --> 00:15:10,020

I want to hear more about Gemini or more about F1, less about sports.

Cue 232 — 00:15:10,060 --> 00:15:14,880

And then it will actually save that to the memory of the agent on the on the

Cue 233 — 00:15:14,940 --> 00:15:19,100

sandbox. And then the next trigger the next morning will then kind of tailor it

Cue 234 — 00:15:19,140 --> 00:15:25,340

again to the to my liking. And so you can schedule triggers to run at any period of

Cue 235 — 00:15:25,380 --> 00:15:30,060

time at any day. And they run in the background.

Cue 236 — 00:15:30,100 --> 00:15:34,620

So essentially you get an ID when it triggers and then it comes,

Cue 237 — 00:15:34,660 --> 00:15:38,380

you can kind of pull that ID and when you come back to it, it has a status.

Cue 238 — 00:15:38,420 --> 00:15:41,660

So it'll say incomplete. And then when it moves to done you can then kind of pull in

Cue 239 — 00:15:41,700 --> 00:15:44,280

the context and actually see what it did.

Cue 240 — 00:15:47,300 --> 00:15:50,660

So like we said, you can kind of schedule these triggers like a morning a morning

Cue 241 — 00:15:50,700 --> 00:15:52,780

issue solver, kind of a QA runner.

Cue 242 — 00:15:52,820 --> 00:15:53,980

These are all ones that we use.

Cue 243 — 00:15:54,020 --> 00:15:57,300

We I use them for GitHub repositories that I don't want to put up,

Cue 244 — 00:15:57,340 --> 00:15:59,360

but I don't necessarily want to maintain.

Cue 245 — 00:16:00,060 --> 00:16:04,500

And I'll have it run to check for new issues, go ahead and reproduce them as well as

Cue 246 — 00:16:04,540 --> 00:16:09,260

kind of show runners. So nightly we'll run benchmarks and see the latency,

Cue 247 — 00:16:09,300 --> 00:16:13,220

see the pass fail rate, and then actually go ahead and suggest changes every day

Cue 248 — 00:16:13,260 --> 00:16:15,520

based on the changes we're putting in our code base.

Cue 249 — 00:16:15,620 --> 00:16:18,160

And manage agents are a really great way to do this.

Cue 250 — 00:16:18,340 --> 00:16:20,820

Also kind of like SRE and security.

Cue 251 — 00:16:20,860 --> 00:16:24,620

So whenever we have like an incident or a new ticket comes in we because it's just

Cue 252 — 00:16:24,660 --> 00:16:28,500

an API call will trigger a managed agent to go do the first pass,

Cue 253 — 00:16:28,540 --> 00:16:32,620

see what the customer is saying, go ahead and actually try to kind of reproduce it

Cue 254 — 00:16:32,660 --> 00:16:36,140

if it's a bug or verify that it's fixed already.

Cue 255 — 00:16:36,180 --> 00:16:40,300

So I think there's lots of different cool use cases where the trigger waking up a

Cue 256 — 00:16:40,340 --> 00:16:44,820

certain events or kind of event based architectures also play nicely with managed

Cue 257 — 00:16:44,860 --> 00:16:45,720

agents.

Cue 258 — 00:16:48,580 --> 00:16:52,460

So this one is kind of the example I just said about kind of having issues raised

Cue 259 — 00:16:52,500 --> 00:16:55,580

while you're sleeping. The managed agent, because it's kind of event based,

Cue 260 — 00:16:55,620 --> 00:17:00,660

can actually trigger with webhooks based off of that kind of issue being opened.

Cue 261 — 00:17:00,700 --> 00:17:04,440

It can go ahead and actually fix the bug, put up a pull request,

Cue 262 — 00:17:04,660 --> 00:17:08,140

review the code, and then you come and wake up and it's actually fixed.

Cue 263 — 00:17:08,180 --> 00:17:12,500

So really cool use case. And then this is kind of the benchmarking.

Cue 264 — 00:17:12,540 --> 00:17:17,540

One can actually test kind of the different test suites overnight and try to speed

Cue 265 — 00:17:17,580 --> 00:17:20,660

them up. We've kind of run managed agents to make sure that our test suites don't

Cue 266 — 00:17:20,700 --> 00:17:25,540

degrade, and that they're hitting a certain benchmark of kind of latency and

Cue 267 — 00:17:25,580 --> 00:17:29,120

quality. And then so we can actually run that every night as well.

Cue 268 — 00:17:32,700 --> 00:17:35,780

And this is another kind of use case I've also been using is this is actually kind

Cue 269 — 00:17:35,820 --> 00:17:38,860

of with all the security issues going on in the world, it feels like there's a new

Cue 270 — 00:17:38,900 --> 00:17:41,800

security incident every single, every single week.

Cue 271 — 00:17:42,020 --> 00:17:46,940

I actually have managed agents check the dependencies on a lot of our repositories

Cue 272 — 00:17:46,980 --> 00:17:51,520

and actually go through and check before we update dependencies,

Cue 273 — 00:17:52,060 --> 00:17:55,740

whether or not they're secure, and go ahead and do some research,

Cue 274 — 00:17:55,780 --> 00:18:00,180

as well as check the kind of make sure they're not vulnerable to any CVEs.

Cue 275 — 00:18:04,740 --> 00:18:06,220

And we can go ahead and run this.

Cue 276 — 00:18:06,260 --> 00:18:10,220

This is an example here where we have like a runner that will run and again manage

Cue 277 — 00:18:10,260 --> 00:18:12,800

agents. Essentially you just see step traces.

Cue 278 — 00:18:13,140 --> 00:18:15,620

And then so if we scroll up here it's going to be really hard to see.

Cue 279 — 00:18:15,660 --> 00:18:16,460

Sorry.

Cue 280 — 00:18:17,620 --> 00:18:19,660

You can see it's executing code.

Cue 281 — 00:18:19,700 --> 00:18:24,140

It's thinking you can go ahead and actually put these into really nice UI

Cue 282 — 00:18:24,180 --> 00:18:27,420

experiences. And then get the files.

Cue 283 — 00:18:27,460 --> 00:18:31,540

We have files APIs. So you can actually take screenshots from the managed agent for

Cue 284 — 00:18:31,580 --> 00:18:36,500

verification purposes. Get JSON, any type of file back and just download it from the

Cue 285 — 00:18:36,540 --> 00:18:39,560

sandbox, or take the whole sandbox as well.

Cue 286 — 00:18:39,700 --> 00:18:42,540

And then we can again run another agent where this is reusing the same exact

Cue 287 — 00:18:42,580 --> 00:18:46,180

environment that we just ran. We're picking up and actually going to go ahead and

Cue 288 — 00:18:46,220 --> 00:18:49,880

make the code fix. So we're editing files, calling different tools,

Cue 289 — 00:18:50,420 --> 00:18:52,640

and then we can trigger this third agent here.

Cue 290 — 00:18:53,700 --> 00:18:56,940

It's now going to actually verify again and make a change.

Cue 291 — 00:18:56,980 --> 00:19:00,780

So this is actually running on one of our SDK libraries.

Cue 292 — 00:19:00,820 --> 00:19:02,660

It went through I ran this yesterday.

Cue 293 — 00:19:02,700 --> 00:19:07,100

It went through and actually found one of the demos we have in our library was not

Cue 294 — 00:19:07,140 --> 00:19:09,080

updated and not working properly.

Cue 295 — 00:19:09,140 --> 00:19:12,940

So using managed agents, it's kind of just a QA engineer to go test it,

Cue 296 — 00:19:12,980 --> 00:19:19,660

fix it. And now this is actually merged into the code base, all from just triggering

Cue 297 — 00:19:19,700 --> 00:19:23,320

single little API calls. It's quite powerful.

Cue 298 — 00:19:24,060 --> 00:19:27,420

If you want to get started today, you can go over to Google AI studio,

Cue 299 — 00:19:27,460 --> 00:19:31,580

where it's easy to just kind of send a prompt and just see how to manage the agent

Cue 300 — 00:19:31,620 --> 00:19:36,300

works. You can go into the playground and just trigger it, create some cool video

Cue 301 — 00:19:36,340 --> 00:19:38,860

games. That's another cool use case I do is when I have ideas,

Cue 302 — 00:19:38,900 --> 00:19:43,660

I'll just send a managed agent to actually go build a game and then play test it and

Cue 303 — 00:19:43,700 --> 00:19:47,180

come back when it's done. And then our docs as well.

Cue 304 — 00:19:47,220 --> 00:19:52,560

Ai.studio managed agents, but I think we have some time for some questions now.

Cue 305 — 00:19:52,780 --> 00:19:54,280

And. Yeah.

Cue 306 — 00:19:55,780 --> 00:19:56,580

Thank you.

Cue 307 — 00:20:02,740 --> 00:20:05,460

Yeah. Can you make the slides available?

Cue 308 — 00:20:05,500 --> 00:20:09,580

Yeah, I'm just thinking actually I'm able to read them, but maybe.

Cue 309 — 00:20:09,620 --> 00:20:12,100

Yeah. Yeah. It's just a single HTML file.

Cue 310 — 00:20:12,140 --> 00:20:14,160

So yes I can definitely put it online.

Cue 311 — 00:20:16,180 --> 00:20:20,140

You mentioned like multi steps for the managed agents and that the agent can

Cue 312 — 00:20:20,180 --> 00:20:25,380

actually get a copy of the environment that the previous agent has been working on

Cue 313 — 00:20:25,420 --> 00:20:30,820

like verify. How do you trigger that like in the multi-step process.

Cue 314 — 00:20:30,860 --> 00:20:36,900

So once the implementation is finished, how do you pass the environment ID or the

Cue 315 — 00:20:36,940 --> 00:20:38,740

agent to invoke involved. Yeah.

Cue 316 — 00:20:38,780 --> 00:20:42,780

Great question. So the question is if you have the kind of the multi-step setup,

Cue 317 — 00:20:42,820 --> 00:20:47,500

how can you actually have one agent kind of when it's done trigger the next one.

Cue 318 — 00:20:47,540 --> 00:20:49,340

And to kind of have that flow go on.

Cue 319 — 00:20:49,380 --> 00:20:54,060

So everything is kind of every interaction has an ID and every environment has an

Cue 320 — 00:20:54,100 --> 00:20:57,780

ID, and at the runtime the agent knows its own ID.

Cue 321 — 00:20:57,820 --> 00:21:00,100

So you can actually right now that's how we do it.

Cue 322 — 00:21:00,140 --> 00:21:04,180

We kind of staged them so that when it's complete, essentially cause another agent

Cue 323 — 00:21:04,220 --> 00:21:07,640

via kind of almost like a tool call saying, hey, handing it over,

Cue 324 — 00:21:07,740 --> 00:21:11,580

we're working on like a better a better flow for this right now through like kind of

Cue 325 — 00:21:11,620 --> 00:21:17,220

sub agents and delegation so that you can actually pass half subjects and say pass

Cue 326 — 00:21:17,260 --> 00:21:20,100

the environment or don't pass the environment off to a sub agent.

Cue 327 — 00:21:20,140 --> 00:21:21,140

So it'll get a little bit better.

Cue 328 — 00:21:21,180 --> 00:21:25,140

But because it's just an API call right now, how I normally do it is kind of just

Cue 329 — 00:21:25,180 --> 00:21:29,500

through almost adding a tool call for the agents to communicate with each other,

Cue 330 — 00:21:29,540 --> 00:21:33,040

each other. So it's built in silico and to be which.

Cue 331 — 00:21:33,140 --> 00:21:36,300

Yeah, not MCP, just you can just do custom function call.

Cue 332 — 00:21:36,340 --> 00:21:38,620

So it's essentially essentially just call.

Cue 333 — 00:21:38,660 --> 00:21:42,720

Yeah. Call this agent. Here's the environment ID of the environment.

Cue 334 — 00:21:43,100 --> 00:21:44,640

And then you get it when it's done.

Cue 335 — 00:21:47,260 --> 00:21:51,860

You talk more about secret storage living on the the boundary there and how all that

Cue 336 — 00:21:51,900 --> 00:21:53,820

works with the classifier. Yeah.

Cue 337 — 00:21:53,860 --> 00:21:58,260

So the question is kind of the more the security aspect and the credentials.

Cue 338 — 00:21:58,300 --> 00:22:03,860

So yeah, our credentials API essentially is so we have agents.create which is kind

Cue 339 — 00:22:03,900 --> 00:22:10,140

of creating. But we also have kind of credentials API for essentially it maps a kind

Cue 340 — 00:22:10,180 --> 00:22:15,980

of like network URL. So if you like for instance, it was github.com and says here's

Cue 341 — 00:22:16,020 --> 00:22:19,380

the bearer token. And then kind of what it does, like I said,

Cue 342 — 00:22:19,420 --> 00:22:22,740

is it puts in a dummy token into the sandbox.

Cue 343 — 00:22:22,780 --> 00:22:24,740

So everything thinks it has everything.

Cue 344 — 00:22:24,780 --> 00:22:28,200

And then when it goes to actually make the call, we swap that out.

Cue 345 — 00:22:28,380 --> 00:22:33,520

If it matches the URL or network requests, that is outgoing from from the proxy.

Cue 346 — 00:22:34,220 --> 00:22:36,580

So you can think of as well as like the MCP.

Cue 347 — 00:22:36,620 --> 00:22:41,860

So when you do MCP auth, you can kind of do auth, you get a URL back and a token and

Cue 348 — 00:22:41,900 --> 00:22:46,300

then you can use it for MCP or yeah, just answer your question.

Cue 349 — 00:22:46,340 --> 00:22:47,880

Or was there a specific question?

Cue 350 — 00:22:48,060 --> 00:22:50,220

Well, I think you were for, for scheduled tasks.

Cue 351 — 00:22:50,260 --> 00:22:54,300

Right. That would require like a user to verify their operations or something like

Cue 352 — 00:22:54,340 --> 00:22:56,100

that to take an action or something.

Cue 353 — 00:22:56,140 --> 00:22:58,900

Yeah. So that's that piece. You have to.

Cue 354 — 00:22:58,940 --> 00:23:03,200

So essentially when you add a credential through the credentials API,

Cue 355 — 00:23:03,380 --> 00:23:05,660

you get like an ID back as well.

Cue 356 — 00:23:05,700 --> 00:23:08,500

And then you can give that to your scheduled trigger saying,

Cue 357 — 00:23:08,540 --> 00:23:10,420

hey, you've already created this.

Cue 358 — 00:23:10,460 --> 00:23:13,040

It has the refresh token. It will stay fresh.

Cue 359 — 00:23:13,380 --> 00:23:17,080

Here's the credential ID and go use it.

Cue 360 — 00:23:17,220 --> 00:23:18,760

But the trick is keeping it fresh.

Cue 361 — 00:23:19,180 --> 00:23:20,600

Yeah. So the credentials.

Cue 362 — 00:23:21,900 --> 00:23:25,480

Yeah. So the credentials API will will will refresh it.

Cue 363 — 00:23:26,740 --> 00:23:29,180

If it has, if it has the ability to do so.

Cue 364 — 00:23:29,220 --> 00:23:35,980

Yeah. When do you like this. Have gone to 150

Cue 365 — 00:23:36,020 --> 00:23:38,420

pages. And we are used to subagents.

Cue 366 — 00:23:38,780 --> 00:23:45,620

And if you use compaction or something like JSON file, what's the method you use for

Cue 367 — 00:23:45,660 --> 00:23:51,200

like the normal like kills the other progress and like for.

Cue 368 — 00:23:52,260 --> 00:23:56,860

Yeah. So so the Antigravity harness that runs in the managed agents handles all of

Cue 369 — 00:23:56,900 --> 00:24:01,740

that for you. It does the compaction and it does the like.

Cue 370 — 00:24:01,780 --> 00:24:05,060

It's the full harness so that you don't have to you don't have to do that.

Cue 371 — 00:24:05,100 --> 00:24:09,980

There's knobs like the max total tokens and compaction limit that you can set so you

Cue 372 — 00:24:10,020 --> 00:24:15,900

can customize it. But we kind of manage that so that it's hopefully you don't have

Cue 373 — 00:24:15,940 --> 00:24:20,180

to build your own harness or build those into your, your.

Cue 374 — 00:24:20,220 --> 00:24:24,120

Workloads when you think about it from Fashion Cell, but only from the fact that I'm

Cue 375 — 00:24:24,160 --> 00:24:28,780

building or you like using in the Dom.

Cue 376 — 00:24:28,820 --> 00:24:33,220

So there's like 150. Yeah. I think our default is 256K.

Cue 377 — 00:24:34,180 --> 00:24:41,020

Like we do a bunch of evals and tests to kind of set a sensible default agents

Cue 378 — 00:24:41,060 --> 00:24:46,560

as they like work, work longer towards that million context token window.

Cue 379 — 00:24:46,820 --> 00:24:50,620

We find like it's harder for them to keep everything straight.

Cue 380 — 00:24:50,660 --> 00:24:52,640

So we find for like most tasks

Cue 381 — 00:24:53,860 --> 00:24:59,080

256K is kind of like the sweet spot, but you can configure that.

Cue 382 — 00:24:59,900 --> 00:25:00,700

Yeah.

Cue 383 — 00:25:02,740 --> 00:25:04,780

Awesome. I think that's it for questions.

Cue 384 — 00:25:04,820 --> 00:25:05,960

Thank you very much.

skills

talk-wotherspoon-agent-execution-environments

tile.json