Hi everyone. Thank you for being here. So today we're going to talk about Codex. My name is Katya Gilusman and I'm with VB. We are both working in the developer experience team at OpenAI based in London. And so our role is really to help developers build and get the most out out of our products including Codex.

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 our 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. For example, GPT 5.3 Codex was our previous ones. We also have the Spark version which is like the super fast model that we have. The state-of-the-art model right now is GPD 5.4. And we also have a mini version that came out last week.

And you know, 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 evaluates the agents' 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, the safety embedded in that harness. So all of that is Codex.

We also released GPD 5.4 mini and GPD 5.4 Nano which you can use for short-running tasks and sub-agents which we'll talk about in a bit. One thing that as we 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 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 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 services like Slack for example. At OpenAI we all the time just like 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 know, you can integrated with Figma, with linear, with notion, all of that combined can let Codex do everything that a software engineer colleague would do.

Plugins bundle a bunch of things together, like skills, apps, integrations, MCP servers, and they bundle that into reusable workflows. Skills are essentially reusable instructions packaged for specific processes. 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. Apps are connections to other services. The tools that you use every day like Notion, Linear, all of that, you can let Codex connect to it. And MCP servers basically expose tools for Codex to just extend its capabilities further and it's tools from external systems.

Automations is one of my favorite things to do with Codex. You can set up automations that run in the background, like a cron job. For example, you can set an automation to run every day at a certain time. 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 an urgent response? Can you also do a summary of all the things that have happened since yesterday on Slack, bucketed per topic?" Another one is connecting Gmail to check if there are emails that I should actually reply to. This is saving me hours per day.

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 Image Gen. Playwright is essentially like a headless browser, a sandbox browser that Codex can just run and use that to see what it's doing. With the interactive version you can actually click things and navigate your app and take screenshots and analyze those screenshots. Image Gen is a great way to just generate visual assets for your apps and games. I was just like "do a platformer game with platforms made of bricks." That's it. And yeah, it generated everything... literally all the sprites, and I didn't have to do any of that.

One thing that you want to be sure of is whatever it is that your agent produces is of the utmost quality. Codex code review is one of the best in the industry right now. You are able to use Codex code review on GitHub. 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. It would contextualize it with everything that is there on the repo itself. A lot of the times Codex code review will find out changes which would have second-order effects which is not limited to just the diff but also to some other modules which you haven't even touched in the pull request itself. 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.

Sub-agents 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. You can define your own custom sub-agents. For each of these you can define what model you want to use, what reasoning effort, what sandbox mode (read-only or write-access). For a review agent or security assignment, you would almost always 100% want to use it in read-only mode.

In this case, it essentially created review slices... 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. Codex automatically decided that this is potentially a complex task, so it automatically kickstarted the plan mode. It's partitioning all of these persona files and then it's going to spawn 20 sub-agents. You can give these sub-agents more capabilities by giving them MCP access. You can give a sub-agent MCP access to Sentry or one to your Linear backlog so that it can read through all the issues and triage them.

Guardian approvals is an experimental feature. All of us were guilty of using "yolo mode," giving unfettered access to your coding agent. Hence we came up with guardian approval. Whenever Codex needs to run a privileged task—remove a directory, run a server, expose a file—Codex will spin up a new sub-agent which will verify whether or not this needs human interruption. This way we hope to reduce the human fatigue.

Hooks allow you to programmatically ask Codex to do a thing based on a particular event (after tool use, start of session, stop session). For long-running tasks, I created a hook for "stop" which runs a Python script "keep going UI"—it asks Codex to do one more pass, run one more validating command, and then stop.

Is there a way to hand off a task to a cloud agent? Yes, definitely. You can select "cloud" directly from the Codex app. 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. We do want to give you the ability to have your own trusted MCP servers run there or the ability to just have SSH agents that you can spawn off a particular task on a VM.