OpenAI Codex Masterclass — Vaibhav Srivastav & Katia Gil Guzman

Hi everyone, thank you for being here. So today we're going to talk about Codex. My name is Katia, Katia Gil Guzman, and I'm with Vibi. We're both working on the Developer Experience team at OpenAI, based in London, and so our role is really to help developers build and get the most out of our products, including Codex.

And so today we're going to start with a quick Codex overview, just so we know how many of you here are using Codex. Can you raise your hands?

Yay!

OK, cool. So

we're not going to stay too long on that overview part. And then we're going to do some demos, so we're going to show you plugins and automations. Vibi is going to talk about subagents and then about the bleeding edge.

So hopefully, for those who already know Codex and use it, you'll learn something that you didn't know about. And then we'll have some time at the end for Q&A, so feel free to ask anything. Also, this is a workshop format, so if you have a pressing question, feel free to ask.

And I see you all have your laptops, so also feel free to kind of follow along with us. We're going to show you how to do some things, and you can try it at the same time. And during the Q&A, that's also the perfect time as well to try things on your side.

OK, so

to start, just for those who don't know Codex, or even if you know it, maybe you don't know it that well, Codex is OpenAI's software engineering agent. So it's not just a coding agent. It's not just

an agent that writes code. It can do much more than that. It can run commands, it can run tests, it can explore code bases. It can really do everything that a software engineer would do. And so it's based on our models as a foundation.

So for example, GPT-5.3 Codex was our previous one. We also have the Spark version, which is the super fast model that we have. The state-of-the-art modelright now is GPT-5.4, and we also have a mini version that came out last week.

And every time we make improvements, every time we have better models, Codex benefits from it. But it's not just the models. On top of that foundation, we have what we call a unified agent harness that will manage, evaluate the agent's behavior, and that is a wrapper for tool execution, for environment setup, for everything that can let the agent do its work and run smoothly.

There's also

safety embedded in that harness. So all of that is Codex, and then you can interact with it through different surfaces. So you have the Codex app that we're going to talk about in a few minutes. You can also interact with it through your IDEs with the extension.

You can interact with it through the CLI, and also through other surfaces like Slack. For example, at OpenAI, we all the time just ping Codex in Slack and ask it to fix things, or in GitHub as well. And on top of all of that, you can also integrate it with your preferred tools so that it can really work with everything that you're already using.

So you can integrate it with Figma, with Linear, with Notion. All of that combined can let you really do everything, can let Codex do everything that a software engineer colleague would do. And so as I mentioned, this is based on our models, and so I'm going to let Vibi tell you a little bit more about that.

Allright. Good morning, everyone. So as we've been talking about the Codex app, the IDE extension, the CLI, and so on and so forth, all of these harnesses, as well as all of these surfaces, would not be nearly as good without the models powering them.

And just to sort of take a step back, back when I joined OpenAI, which is not really as far back along, was in December, our leading model at that time was GPT-5.2. And

from there, we went on to release GPT-5.2 Codex, which was a specialized Codex variant of GPT-5.2, wherein we sort of pushed how far you can take the model and run it on long-running tasks, how far you can let it just continue to chug along.

And then shortly after, we followed up with GPT-5.3 Codex. Shortly after that, in partnership with Cerebras, we followed up with GPT-5.3 Codex Spark. And most recently, we released GPT-5.4. And you can already see how we're sort of pushing this whole sort of model and harness flywheel as fast as possible, trying to bring the next frontier as fast as possible to you all.

Something to note is, and something that's not on the screen, is at the same time, whilst we were pushing for larger models, which are really good for long-running tasks, as well as really complex tasks, and so on and so forth, we also released GPT-5.4 mini and GPT-5.4 nano, which you can use for short-running tasks and subagents, which we'll talk about in a bit.

And something that we haven't really emphasized on this over here is two things. One, that as we sort of pushed on making these models better, we also worked quite a bit on making sure that these models can be served to you as fast as possible.

What that means, in principle, is we introduced something called WebSockets, which allow us to sort of create a connection between your device, as well as where the API resides, to be able to give you roughly about 1.75x faster tokens

without really paying the cost over to you. At the same time, we also released a fast mode, which allows you to, on top of the 1.75x, get 2x more faster tokens. And this is something which the team is continuously sort of hammering on.

There's lots more speed improvements coming over there.

And so to bring this all together, at the start of this year, we brought together the Codex app. How many of you have used the Codex app? Allright, that's a fair good chunk of people. To be honest, back in December, and even before, I was a hardcore CLI user.

And at some point during the app launch, whilst I was beta testing it, dogfooding it, the Codex app became a really important part of my workflow. And the reason for that is it brings together a really nice way to work across projects, number one.

And number two, within a same project, work on multiple features at the same time. The way you can do that is you can have individual projects, like you can see on the left side, you can have the Codex project, ChatGPT, Sora, and so on and so forth.

But also within those, you can use work trees to work on individual feature requests, or bug fixes, or just Q&A, all at the same time, without really interfering with individual tasks. This is something which we're quite proud of.

And providing a native work tree support helps you do the same task and do multiple tasks at the same time without really having to context switch as much. At the same time, through the launch, we've been trying to sort of increase the net benefit you can get out of the Codex app.

And some of these features have been like having a better automation support. Automations is also something we're going to talk about just in a bit. But the short summary of automations is that you can essentially have a rough process that you want Codex to run, let's say, every day at 9:00 a.m., or let's say, every evening, or let's say, you want Codex to look through your calendar and create a briefing for you.

And that all is possible, all within the native Codex app setup with automations. And then, of course, with the work trees and more native Git support, you can sort of work across projects and just be able to push changes as you want with whatever Git persona you want to do it with.

Last but not the least, based on which surface you use the Codex app on, at the start of the year, we released it just on macOS, but now we have native Windows support, which comes along with native Windows sandbox.

Is there anyone here who's using Windows today? I'm cheering for you, man. And so for the one gentleman over there, we have native sandbox support in Windows. We're the first of kind. There is no other competing harness which supports a native sandbox for Windows.

Cool. And so I've been talking on and on about the Codex app, been talking about all the models that we've been shipping. But what's new in terms of all the features that we've shipped?

This is in, I think, if I'm not wrong, in the sort of descending order. So most recently, we launched Plugins. Plugins is a way that you can bring together skills, MCPs, as well as prompts, and any other thing really, together in one bundle and allow the model to do more nuanced matching whilst it's building.

We also released recently mini models, which tie in quite well with subagents, which allow you to parallelize a particular feature, or bug request, or Q&A, whatever it may be,

at a faster rate, all whilst making sure that you don't pay as much cost for your particular models. And then we have a bunch of other stuff which we're going to talk about as we go through. Some of this is how Codex is so good at code review, how Codex is really good at security, and so on.

All of this, whilst we talk about all of this, I want to sort of emphasize on this fact that we're at OpenAI quite lucky that the community has really embraced Codex. In fact, just last night, we crossed the milestone of crossing 3 million weekly active users.

And this is a pretty big deal for us. And we want to continue sort of supporting the developer community, the enterprises, startups building on Codex. So throughout this session, if you have any questions, please feel free to throw it at both Katia and myself, or even afterwards, or just ping us.

With this, I'll pass it over to Katia.

Thank you. And yeah, the 3 million weekly active users thing,

it's really cool to see. And it's crazy to think that it's also more than tripled since January. So just in a few months, we've seen huge adoption. And yeah, it's really, really cool to see. OK, so Plugins. Plugins, I don't know if you've heard about it.

It's quite new on Codex, the native support for plugins. The idea of plugins I'm going to show you what it looks like in practice and how you can use them is that they bundle a bunch of things together.

So like skills, apps, integrations, MCP servers, and they bundle that into reusable workflows. And so what skills, apps, and MCP servers are, again, I'm going to show you. But just to introduce that a little bit, so skills are essentially reusable instructions packaged for specific processes.

So if you have something that you're doing quite a bit, you can actually create a skill for it so that Codex knows about it. You can give it instructions. You can give it scripts as well, resources. And all of that will save you from just repeating yourself over and over.

So every time you have a sort of neat workflow that is always the same, you can package that into a skill. You can actually ask Codex to create the skill for you as well. And then apps are connections to other services.

So again, we'll see a quick demo. But the tools that you use every day, like Notion, Linear, all of that, you can let Codex connect to it. And MCP servers, you might be familiar with this already, but they basically expose tools for Codex to just extend its capabilities further, and it's tools from external systems.

And so all of these three things are already very useful on their own. And what plugins do is that they bundle that so that you don't have to set up everything manually. You don't have to install multiple skills.

You don't have to connect multiple apps. You don't have to connect multiple MCP servers. You can just add a plugin. And another thing I wanted to talk about in the Codex app, and that we'll show a quick demo for, is automations.

Personally, this is one of my favorite things to do with Codex, because you can set up automations that run in the background, so like a cron job. And you can connect apps. You can use plugins there, too, and just set it to run on a scheduled time.

So for example, you can set an automation to run every day at a certain time. And it's just an instruction that Codex will run in the background.

And the last thing I wanted to show you with the demoright after is specific skills for web, app, and game development, because

we've heard a lot about developers who want to use Codex to build these things, to build apps, to build games. And every time, they kind of repeat themselves. Every time, they kind of use the same skills. So we actually package that into specific plugins.

And there's two skills that I want to highlight that are super useful and, honestly, that are a game changer when you're developing something visual. It's Playwright Interactive. And so for those who don't know, Playwright is essentially like

a headless browser, like a sandbox browser, that Codex can just run and use that to see what it's doing. So you can open your app in the browser. And with the interactive version, you can actually click things and just navigate your app, and take screenshots, and see and analyze those screenshots.

And then Image Gen is a great way to just generate visual assets for your apps and games. So enough talking. I'll show you a demo. I'm going to start by actually running this one, because this one is pretty long.

When I ran it yesterday, it took like an hour to build. So I also have the final version. But I wanted to show you this prompt, how Codex is going through it. And so what I'm doing here is I'm using the Game studio plugin, which is, again, a bundle of a bunch of skills that are helpful for game development.

And I'm asking it to use Image Gen to create visual assets, so sprites, for the games, and using Playwright Interactive to also debug the game and make sure that it works well. So we're going to let that run.

And then we're going to talk about plugins for a little bit. So let me switch to another project here,

so this developer's website one. OK, so this one is the repo for our developer's website, which is

here. Sorry, I'm going to put that in full screen. And so on our developer's website, we have this page with all of the Codex meetups we have. So there's a lot. And all of that is actually in our repo, like in our code base in YAML files.

And so what I'm going to do is I actually added this Google Drive plugin here. We have a lot of featured plugins built by us that you can choose from. You can also, of course, add your own plugins.

But I connected this Google Drive plugin that lets Codex access my Google Drive. And so what I did is that I prepared this spreadsheet called Codex events with the event name, date, and city. And I'm going to ask Codex to just update this sheet with the current Codex meetups listed in the code base.

So I'm going to start this. Again, it's going to take a little while. And so let's check in on it. OK, for the game task, it's still running. I'm going to show you when it's doing a little bit some more interesting things.

But the last thing that I mentioned is automations. And so automations is, again, something that you can just set up using apps. You can just ask Codex anything. But instead of it being interactive, like you're actually using the Codex app, you can set it up to run in the background.

So for example, some of the automations that I set up that are honestly helping me a ton in my day-to-day lives is one for Slack messages. So I connected Codex to Slack. And I'm asking, hey, Codex, can you check every day at 9:00 a.m.

the messages that I should reply to and flag if it's time sensitive or waiting for anurgent response? Can you also do a summary of all the things that have happened since yesterday on Slack? And I'm asking that to bucket it per topic and then important information to be aware of.

So we have important channels where company information, generally the things

that leak in one day. So

important company information is in there. And so I just want to make sure that I don't miss anything here. So that's the kind of stuff that I ask Codex to just summarize for me. Another one that is pretty cool is connecting Gmail.

And same thing, I receive honestly an ungodly amount of emails per day. And so I'm just asking Codex to check if there are emails that I should actually reply to and to check if it's time sensitive or if it looks legit or not, because I do get a lot of requests that are not necessarily something that I would reply to.

But this is saving me hours per day. And so the way you can create automations is you can create it from here, or you can also just

say something like,

hey, Codex, can you create an automation that will look at Slack and look for anything that mentions Codex use cases and then list all of the important use cases that

I should put on our website?

So I'm going to let Codex think about this for a second. I should have used Spark.

And it's going to come up with this. It's going to create the automation for me, basically. And I didn't specify when I wanted to run it. But I can actually like, oh,

interesting. It's doing something different, because this is a live demo. So obviously, it wouldn't OK, normally, it should do a little pop-up. So I can just click on the oh, it's doing it. Perfect. It was just very chatty this morning.

OK, interesting.

Interesting. OK, so please create the automation.

So it should show a little pop-up if everything goes well. But if not, you can still create it manually. Let's just see if it is doing it. OK, I don't know what's going on. But OK, let's just do it manually.

So

you can also create it from here. And basically, all you have to do is just call the plugins you want to use, like use Slack, and then choose the frequency where the automation should run, which project it should run in, et cetera.

OK, so let's check on our other tasks.

This one is still running. OK, it generated some pretty cool sprites. We'll look at this after. Let's check on our task to update the spreadsheet. So here, Codex took two minutes to actually analyze the code base. It found the source for all of the Codex events where we have our YAML files.

And then it wrote the 57 event rows. So we have 57 events currently listed on the website. And so let's check. Let's see our spreadsheet. And yeah, we can see that it was updated. Nice. So this is something this is a simple example.

But every time you have something that's very time-consuming and anything that has anything to do with data, data review, for example, you can actually ask Codex to do it for you. It has access to everything on your code base.

And you can also feed it other inputs, like other CSV files. And then you can just ask Codex to do that type of work for you. OK, now, last thing, let's check on our

game. So as you can see, Codex is actually using Image Gen to generate I'm going to zoom out a little bit. So oh, nice. So it's generating all the sprites, all the game assets that I asked it to do.

And this looks pretty nice.

So it's going to take a while. What I'm going to do is I'm actually going to show you final results. But as you can see, Codex is just reading sorry, it's just generating all of these assets. And then it's going to use the Playwright skill to see how that looks like in the app.

So unfortunately, we don't have an hour to wait for this final result. So let me just show you the one that it did yesterday. So this is

untouched. I haven't touched it. It's literally just Codex who built this. And all of that I gave zero input. I was just like, do a platformer game with platforms made of bricks. That's it. And yeah, it generated everything.

So granted, the overall UI is not like I would probably iterate on that. But

I think the platformer itself is pretty cool. And what is really cool here is that literally all the sprites like here, I'm just moving all around. And that's at least like five different sprites of the little character. And I didn't have to do any of that.

You can also do a custom game with your face as input and have Image Gen just create a 2D version of you. So that's a way that you can leverage the Image Gen skill, the Playwright Interactive skills, and that Game studio plugin.

And just to show you what's inside, we have also the same thing for web apps. But it's a bundle of all of these skills together. So yeah, that's it for me. I'm going to pass it back to VB.

Thank you.

Thank you, Katia.

Allright,

perfect. So just to do a very quick checkpoint and a recap on what we've spoken so far, so we went through all the models that power the Codex ecosystem. Then we went through all the surfaces you can consume Codex from.

And then we went through plugins, how to use them, and what are some of the plugins that you can use. You can also create your own plugins using Plugin Creator. And then we went through to speak about automations and Image Gen and so on and so forth.

Now, something to note is as we continue sort of delegating more and more work on these agents, it could be any of your favorite agents, Codex or not, one thing that you want to be sure of is whatever it is that your agent produces is of the utmost quality, which means that as we start sort of working on multiple features at the same time, multiple projects at the same time, it's going to be quite likely that it's impossible for you to go and look through each and every line of code, which means that at least for the first pass, you want to have a way which you can rely on to review your code.

And this is where code review sort of comes in.

By no means am I bragging about this. But in my own biased way, Codex code review is one of the best in the industryright now. This is something which people on Twitter and LinkedIn, on our own sort of platforms, Discord, and so on and so forth, keep raving about that how is Codex code review so good.

So I wanted to spend a quick, hot minute on what it does. So first of all, it is available on the surfaces that you work at, which means, number one, you are able to use Codex code review on GitHub.

So you can connect your ChatGPT account with GitHub. And for each and every pull request that you create, you can set it up such that Codex can automatically review each and every pull request. And it would typically give you

some sort of a

what's this called? A callout like this on the pull request itself saying that, hey, this is something that is missing. Hey, maybe P0, fix this. P1, fix that. P2, this is something that would be good to have, and so on and so forth.

At the same time, you can use /review on the Codex CLI or the Codex app. And Codex will spin up large sort of review process and so on and so forth. And very recently, last week, with my colleague Dom, we shipped a Cloud Code plugin for Codex, which allows you to essentially invoke Codex within your Cloud Code sessions to be able to get the same sort of state-of-the-art code review, but in your Cloud Code sessions.

So something to sort of see here is let's say that I am working on a project like this. By the way, this is my actual working setup at work. This is all which I work on. Everything that you see here is all of these threads, all of these projects is something which I work on day to day.

So if you see something which you shouldn't, just close your eyes. And so typically, what I would do is I would go through a feature request, or I would go through some sort of ask from someone. And let's say over here, I ask Codex to do a bunch of things.

So I'm just going to ask it to review its changes.

And so then you get an option to either choose from a base branch. If you have multiple branches in the Git repo, you can choose it against a feature branch, against an eval branch, whatever it may be, and so on and so forth.

In this case, I'm just going to ask it to review uncommitted changes. And what it does is if you see

here, what it does is it spins off a totally new thread. And what that thread would do is it would essentially spin up a totally new Codex process, which has our own review system prompt. And it would continue sort of looking through not just the diff or the list of all the changes, but it would also contextualize it with everything that is there on the model repo itself.

And so a lot of the times, Codex code review will find out changes which would have second-order effects, which are not limited to just the diff or whatever changes you've made, but also to some other modules which you haven't even touched in the pull request itself or in the changes itself.

And this is so effective that 100% of pull requests across all OpenAI repos made by all employees, including Greg, are reviewed by Codex code review by default. And that's the first pass that you take.

Cool. And so as you can see over here, Codex worked for a minute. And it came up with these sort of updates like P1, localize whatever revenue detail. P2, translate this to this, and so on and so forth.

And what you can do after this is essentially ask Codex to either take a pass at fixing this or open another sort of PR on whichever branch you're at and then sort of go on from there.

Now we get to subagents, which is something which I'm personally quite excited about. So first and foremost, what is subagents? Subagents is essentially the ability wherein you can spin off a master task into decomposable, parallel, and independent tasks, which you can hand off to agents, which can allow these agents to sort of work independently and then, at the end of their run, get back to you and give you a response.

And over here, sky is literally the limit. You can spin up as many agents as you want, of course, as long as your API key or your whatever ChatGPT Pro Plus Go subscription you're on can take. You can do a lot of interesting things with subagents.

For example, what I'm doing on the screenshot on the left is I have a Codex agents repo, which we're going to look at in a sec. It's not public yet. But I hope that we'll be able to make it public very soon, which has a lot of personas for subagents that you can use.

So it's kind of meta. It's essentially subagent personas like doc reviewers or test case creator or test case runner and so on and so forth. And

every now and then, we would change the spec. This is from before we wanted to change the spec of how subagents work. So what I wanted it to do is to go through all of these 40, 50 different subagent personas, review them, and make sure that they are up to spec.

And of course, doing it without subagents would have meant that Codex would open each and every file and then review it and then give me a summary and continue doing it for like 50 different subagents. In this case, it essentially created review slices, which means it created, say, these are the two files that

subagent Polly or subagent Plato should essentially review. And then they would spin up a new Codex environment. They would review those. And then at the end, Codex will collate all of these and give me back a response. So let's give this a shot.

So the repo in question is this. It's just the Codex agents repo, which has a bunch of personas. You can see that we have quite a few sort of personas over here. We've got an accessibility reviewer, architect, and so on and so forth.

And this is actually something which you can create yourself. And we're going to touch on that in just a minute. You can define your own custom subagents. But think of this repo as a collection of these subagents. And this is typically what you would have for each and every subagent.

You would have a name. You would have a description. You would have a different sort of sandbox mode, whether you want it to be write only, whether you want it to be read only. And then you would have some sort of instructions and so on.

And so now what I'm going to do is I'm going to ask Codex to

I'm going to go over to my Codex agents. I'm going to switch to let's do medium over here. Let's close this.

Can I make this full screen?

Allright. So let's give it a task. Spin up 20 subagents to review all the subagents.

So this is a very simple task. All I'm asking Codex is to do the same task which I was showing before, wherein I wanted to review all the different subagent personas in this repo. And you can see that

it already figured out that there's agents and skills. And it's looking into it. There are 45 curated persona files. And what it's going to do is it's going to create 20 reviewers. And

it's going to give them all of those

Tamil files. And then it's going to review those. And you can see that there's two things which are quite interesting over here. Number one, Codex automatically decided that this is potentially a complex task. So it automatically kickstarted the plan mode, which is what's active over here.

So you can see that it essentially came up with five tasks to solve this particular problem. You can explicitly invoke plan mode as well. But in this case, it decided to do it on its own. It's then partitioning all of these persona files.

And then it's going to spawn 20 subagents very soon.

I swear it's faster.

So now what it's doing is it's

oh.

So for some reason, on my particular setup, I have a cap on six concurrent agent threads that can be run at the same time. We can fix that. But to go back up, what we can see over here is that it actually spins up six agents, which is my limit for now.

And I can see all of those agents working over here. I can quickly see what Jason, the agent over here, is doing or Hume and so on and so forth. And you can see that something to note here is that the main Codex model over here hi.

The main Codex model over here essentially created a persona. And not just that, it doubled down. And it gave the exact files that this particular subagent should review.

And additionally, it also gave it some insight on there's repo guidance in repo.md, in contributing.md, in skills, and so on and so forth. And it will sort of continue going down this route for all the different subagents. And what it does towards the end stage is that it will tear down all of these subagents

when they have gone through when they have gone through their whole process of looking through all the Tamil files and so on. And if I go back to my main thread, you can see that two of the agents are still working.

But eventually, it would collate all of this feedback that it has gotten from all of these individual subagents and proceed. Now you can think of this this is like a very simple sort of explorer use case. But you can think of this from, for example, a cybersecurity perspective, wherein you have a Git commit or you have a particular Git repo.

And you want Codex to spin up and run multiple vulnerability

sorry, one sec. You want it to create multiple sort of vulnerability analysis from different points of views or from different hypotheses. And you want it to sort of tackle the same diff or the same GitHub repo and try and come up with a vulnerability map.

And this is something we actually use

quite a bit or I personally use quite a bit when I'm brainstorming a particular feature. I would just spin up multiple Codex subagents to sort of look through how I would approach a problem. So let's say I want to add a feature.

I would ask Codex to create a plan for what are, say, five or six or ten different ways that

a particular feature could be implemented. And then I would quickly double down and ask Codex to then create multiple subagents to get me some sort of understanding for these tasks. Sorry, my watch was constantly vibrating. And so that's like a quick high-level overview of how subagents work.

By default, we ship three subagents, three subagents personas.

Let me quickly open.

So by default, we ship three personas. One is a default general-purpose fallback agent. Another is a worker, which is sort of execution-focused. So this is something that you would use

when you want Codex to write a particular feature request or work on a particular feature. Then there's explorer, which is the same one which we used before. And then for each of these, you can double down and create your own Codex subagent personas like we saw before.

And we will create oneright now. Something to note is that these particular subagents,

for each of these, you can define what model you want to use. You can define what reasoning effort do you want to use. You can define what sandbox mode do you want to use and so on and so forth.

The reason why this is important is for a review agent, you would almost always 100% want to use the review agent in read-only mode. You would never want your review agent to execute anything. For the same reason, for a cybersecurity vulnerability assignment, you would want your subagent to always be in read-only mode.

But

for a docs writer or for something which creates docs for a particular feature that you've created or a bug report and so on, you do want to give it write access so that it can execute stuff and also

create a bug report for it as well. Something to note is that you can also double down and give these subagents more capabilities by giving them MCP access. So you can just give let's say you can give a subagent MCP access to Sentry so that it can look through

all of your reports over there or one subagent access to your linear backlog so that it can interact with linear. It can read through all the issues added to you, triage them, and so on and so forth. You can also give them skills.

So really, if you really want to, you can quite heavily customize this entire setup

for your own

use case.

So let's open our Codex app again. You can see that it went through all of these subagents. It created a bunch of other subagents just to go through all of these. And it came up with these findings. It's based on README, based on contributing.

Performance investigator is overprivileged. P1 has a sandbox sorry, verifier has a sandbox mismatch. Same for writer and so on and so forth. And so you can see this is already quite useful. And it saves you quite a bit of time to be able to go through all of these individually or sequentially and so on and so forth.

Now let's go back and see a bit more about custom subagents.

So as I mentioned, we ship three subagent personas. But at the same time, you can create your own custom subagents. In fact, we do recommend creating your own subagents or just ask your Codex to look through your past sessions and create subagents for you.

Both of these scenarios work and work quite well. So in this particular case, you can see that we have a PR explorer subagent, which reads your code base, uses GPT-5.3 Codex Spark, which is our research preview model text-only deployed on Cerebras and is blazingly fast, is quite fit for this particular use case.

And we set sandbox to read-only. So we don't want the model to sort of execute. And we give it certain

instructions. So in this case, we say, hey, stay in the exploration mode. Trace the execution path.

Don't propose any fixes. And just search through and figure out what exactly do you want us to do.

Now let's quickly

try and create a subagent. So let's say we want to do docs researcher. In this case, what I typically do is to just go and ask,

hey, Codex, can you create this subagent for me? Here's its persona. And then let's see.

And so what Codex is going to do because Codex is aware about how it works and what it's supposed to do and where it's supposed to place all of these things. What it's going to do is it's going to create a Tamil file for this docs reviewer.

And in this particular case, this uses the docs MCP server, which we created

from the DX theme, which packages all the API references, all the docs, all the guides, all the toolkits, and so on and so forth. And it will add that as an MCP server so that every time we ask it a question about, hey, what's the best way to use GPT-5.4 with WebSockets?

Or what's the best way to use GPT real-time

with I don't know, pick your favorite way of using GPT real-time. And can you create a React plugin for this and so on and so forth? It would be able to reference all of these things. So I'm going to let it do its thing.

And in the meantime, head back over to the slides.

And so just to go back sorry, one second.

What you can do just to sort of invoke

a particular subagent is you can say, hey, can you

reviewer

subagent and review each and every persona based on the developers' docs. So in this case, you can essentially use the same particular subagent, leverage it again, and then ask it to do the particular task that you want to do.

Now what are some interesting ways that you can use this is imagine you have a long build process or you have a test process. You can have a subagent which can run your test case locally. You can have a subagent which can always make sure to oh, I'm being told that I don't have as much time.

You can have a subagent which can pull the latest from GitHub as soon as you do a pull. You can have a subagent which can quickly pull all of the context from a linear issue and so on and so forth.

So really, you can do this for you can leverage this for a lot of things. And the best thing that I like to do is to just ask Codex to look through my past sessions and recommend me certain automations, certain subagents, and so on and so forth that I can use.

Cool. So now we're at the bleeding edge. This is a bunch of stuff which we have shipped in the past. And we haven't really made as much of a splash about. So what we're going to do is we're just going to quickly go around and see what each and every one of these do and how you can leverage them.

So first and foremost is Guardian approvals. This is an experimental feature. You can activate it today by just going on /experimental. So it would be something like

Codex. Hopefully, it works.

And then you can look at experimental. And you can in my case, I already use Guardian approvals. And you can activate it this way. What Guardian approval does is all of us, including myself, at some point, were guilty of using YOLO mode all the time, which means that you by default give unfettered access to your coding agent to do literally whatever the hell it wants.

And this, by all means and measure, is not safe. Hence, we came up with something called Guardian approval, which for each and every time, Codex needs a privilege needs to run a privileged task. Let's say it is, can I remove this particular directory?

Can I run a server? Can I expose a particular file to the internet? Whenever all of these things sort of pop up, what Codex will do is it will spin up a new subagent, which will, based on a particular prompt, try and verify whether or not this is something which needs my human interruption or not.

And in most cases, it doesn't need human interruption. So it will just say, hey, go on. Run this particular privileged tool or privileged task and so on and so forth. And this way, what we hope to do is we hope to reduce the human fatigue that comes by just always sort of having to approve, do this task, run this particular Bash script, or run this and so on and so forth.

In principle, how would that look is

trying to see if there was

OK, it doesn't show it to meright now. But if I just in the interest of time, I'm going to ask, hey, can you run the dev server? And I'm going to instead of full access mode, which for some reason, again, I'm not able to

click on. Let's try and see if it invokes Guardian approvals. Whilst this works, I'm going to head over to the next step, which is hooks. Hooks is also something which is experimentalright now. We're trying 24/7 to try and make this a better experience.

Currently, Codex supports three hooks. One is after each tool use. One is at the start of a session. And third is when you stop a session. What hooks allow you to do is it allows you to programmatically ask Codex to do a thing X based on a particular event.

So let's say that when you start your Codex session, you want Codex to pull the latest from your GitHub repo. So in that particular case, you would want to set up a start hook. If you want Codex to do something after each tool use, let's say for a lot of researchers who want to document each and every tool use, they might have a per-tool-use hook wherein they document what Codex has done per session and so on and so forth.

So you can do that with that. And last but not the least, something which I personally use is the stop hook, which is when I'm running long-running tasks, at the end of each turn of Codex, I would ask it to keep going so that it just continuously keeps running a particular task.

And

in theory, how this would look like

is where is it?

Sorry, one second.

Wow, I was really prepared for having more time, I have to say. But in theory, how this would look like is that you have some sort of a Python script. And you define a hooks.json. So in this particular case, you can see over here that you have a pre-tool-use.

You have some sort of a matcher. You say, on startup or resume, run this particular session.session_start.py and so on. And you can define how you want to in this particular case. So what I did for the sales dashboard example that I've been showing you so far is I created a hook for stop, which runs this Python script, which is keepgoing.py, which is every time it encounters the stop

hook, it would just ask Codex to keep going, do one more pass, run one solid validating command, tighten one more thing, and then stop and give the result. And so for really long-running tasks, you can just set it up and ask it to continue doing its own thing.

Last but not the least, we have personality changes, which means that you can go on Codex and you can ask it to

quickly look at personalization. You can set up different personalities. You can set up a more friendly personality or a pragmatic personality based on whatever you want to do. You can also add custom instructions. So you can ask it to always cite whatever it is it is doing and so on.

Right, and then last two things

is we released something called Codex Security. This is our state-of-the-art model, which allows you to find and fix vulnerabilities in your GitHub projects. And essentially, what it does is it would go through commit by commit. And it would create a vulnerability patch.

And it would use Codex to then sort of patch the said changes as well.

Lastly, as I mentioned before, we released a Cloud Code plugin, which allows you to use Codex in Cloud Code. This is something which was surprisingly used quite a bit by the community. And this is something which allows you to sort of ask Codex to review whatever it is that you've done so far, run an adversarial review, or just ask Codex to rescue whatever changes you've done so far as well.

That's it. Thank you so much for joining us. And feel free to ask any questions that you might have. So we don't have a lot of time for Q&A, unfortunately. We should have started maybe a little bit earlier.

But happy to take maybe a couple of questions in the room. And then we'll stay here anyway. So if you have questions and you don't have anywhere to leave, you can come to us, yeah.

Thank you so much. I have a question. You said a couple of times that there's a way to scan that Codex scan past sessions and basically give your recommendations. How exactly do I do that? Is that a project with 20 threads or something like that?

How do I scan that?

Yeah, so what you typically do is all of the sessions within Codex are put in .sessions within the same .codex folder. And Codex has the ability to just scan through all your sessions and then do things.

OK, I can do this using the CLI, but not.

Yeah.

I can do this using the CLI, but not using the Codex.

You can use it. You can use Codex app. You can use Codex CLI, anything. You just have to ask it to look through the sessions and do whatever you want to do.

Thank you.

Thanks.

There's another oh, OK, maybe a couple more. Yeah, in the back here.

Hi.

Hi.

Is there a way to hand off a task to a Cloud agent? So let's say I'm here working on a task. And I have to close my laptop. So I hand it off to a Cloud agent.

Yes, definitely. We didn't touch on that. But actually, you can do that from the Codex app directly. Maybe you can share your screen. But you can either work locally. And as you mentioned, you can do it we support Git work trees as well.

But you can also just select Cloud here. And you can select the number of times this task should run. You can parallelize. We call that best of n. So you can run it four times in the Cloud and then just pick the best output.

So that's something that is built in the Codex app in the IDE extension. And you can also access it directly from the web interface.

And there's more cool stuff coming on that very soon.

More what?

There's more cool stuff coming on that very, very soon.

I think there was oneright here. Yeah.

Thank you so much. My question was actually about the Cloud UI as well. Because today, subagents aren't supported, if I'm not wrong. And especially, the thing that bothers me is it doesn't use the skills that are in the repo.

Is that coming soon or?

So there's like a at the risk of

talking about the whole roadmap, we definitely have a lot more changes coming up on that particular front. I'm not sure if skills within Cloud is going to be as soon as I say that it is going to be.

But it's definitely at the top of the mind. And we do want to sort of give you the ability to sort of have your own trusted MCP servers to be able to run there or CLIs and so on, and also the ability to just have SSH agents that you can just spawn off a particular task to on a VM and so on.

So lots of work on that.

But it can use skills in the repo,right? That is checked in.

It's not on Cloud tasks.

Yeah. But if you it reads instructions and stuff. And it can find it and still see it since it's in the code base. It's more like the skills that you have locally that were the same.

The reason why we don't allow it on Cloud is because there's no way for the sandbox to know whether or not a skill is trusted or not. And so that's why we don't. And skill can package like a Python script or an executive script.

It won't execute things. But if you have things like resources, it can access it technically because it is in the repo. It's just, yeah, it's not as good.

So I have to request it.

Yeah.

Thank you.

Thank you.

Were there any other questions?

Cool. Have a great day. Enjoy the day. And if you have any other questions, we're going to be around today, tomorrow, and also maybe on Friday. Feel free to reach out or just drop a DM. And enjoy.

Thank you.