Codex Is a Thought Experiment: What If the Agent Worked on Its Own Computer?

Hanson Wang helped train the Codex-1 model. Alexander Embiricos, Alex, is the product lead. For Hanson the original 2021 Codex was the aha: GPT-3 was cool, Codex was the first time the tech felt world-changing. He was in Sequoia’s Arc program, which is how he met Lauren Reeder, and one of those demos used the old Codex for data analysis. That model powered early GitHub Copilot. The brand then sat unused. Working on the new agent, the team decided the name was too apt not to bring back: code, Codex, code execution. Resuscitate, Alex says. They had not used the brand recently.

The new Codex is a coding agent with its own container and its own terminal, fully in the cloud. You give it a task; it comes back, one-shot, with a pull request. They tried other form factors and settled on this. In Alex’s framing it is a thought experiment. Instead of AI sitting on your laptop pairing with you, put the effort into what it feels like when the AI works independently on its own computer. You delegate. Two pieces they are proud of: a compute environment the agent can actually be productive in alone, and a model that writes code professional engineers can merge, ideally without touching their own machines.

Codex CLI versus Codex in ChatGPT is the question they keep getting. Codex is the brand for agentic coding. The vision is an agent that mostly works on its own computer but can meet you in the terminal, the IDE, the issue tracker. CLI is Codex in your environment. ChatGPT Codex is Codex on its own. Today those are distinct. A brief aside Alex loves about OpenAI: they cut scope and launch. Over time the two will come closer, so it is just Codex, in ChatGPT or in your CLI.

Making the model useful beyond the next line of code meant leaving the contest circuit. Hanson used to compete; o1 and then o3 are better at competitive programming than he is, better than almost everyone at OpenAI. They still were not that good at producing mergeable code. The o3 blog post said as much: generated code often is not to a professional engineer’s taste or style. A lot of Codex-1 training was specialized work to align the model with those preferences. Alex’s product analogy is a precocious competitive-programmer graduate who has not yet had years on a team. The jump from o3 to Codex-1 is those first years: what a good PR description looks like, titles, reading a codebase’s style and matching it, testing well, showing that you tested.

The aha for new users, Hanson says, is often the onboarding task: find and fix a bug. Codex can try to reproduce, not just stare at something that looks off. Leading up to launch they kept pasting issue descriptions in and being surprised how often a usable fix came back. The story that stuck is 1 a.m. the morning of launch. A Lottie animation was broken. They could have cut it. An engineer described the bug, a pro tip if the task is hard: give Codex multiple cracks. Four rollouts. Three failed. One was the fix they had been stuck on for hours. They landed it, deployed, and the animation shipped.

Under the hood, Hanson later confirms, Codex-1 is the same model as o3 with additional reinforcement fine-tuning. Qualitative taste is one half. The other hard problem was environments. Real repositories are varied and messy; standing one up is a DevOps job. Alex walked Hanson through the repo of the startup OpenAI had acquired, thinking of it as a training environment. Hanson asked where the unit tests were. Alex said it was a real startup: there were none. They had to generate realistic environments for the agent to learn in. The reason the end-to-end product works, Hanson says, is that training environments and the container infrastructure that serves users are the same. Sonya Huang’s joke writes itself: so the agent does not get to say it works on my machine. Hanson: exactly.

Internally, the other magic moment is a mindset change. Copilot-style tools pair with you in flow. Codex CLI can still do that. Cloud Codex is built for a future where the vast majority of coding is done independently of the human, who can only do one thing at a time, by agents on their own computers. Delegating is a different skill from pairing. In the alpha they handed people the agent and said use it however you want. Many did not find it useful. Then they looked at OpenAI internal usage and saw the gap: abundance. Try anything, even multiple times, because it saves you time. Onboarding now tries to manufacture that aha by running many tasks in parallel. If someone has run something like 20 tasks in a day, or even an hour, they have probably understood the tool.

That shifts the human job toward review. The team invested in making outputs easy to check. Hanson is proud that the model can cite its own work, not only changed files but terminal commands and outputs, including tests that failed. One of the hosts asks the obvious: if it compiles and does the thing, do you still need a human? For the foreseeable future, yes. Early users have to build a feeling for what works. There is also always external context that decides whether the code is actually correct, beyond what you put in the prompt. Alex decomposes the job: ideate, design, plan, implement, validate, deploy, maintain. Codex is great at the small implement-and-test loop, and you can use it for planning too. A stat he remembers: engineers spend maybe 35% of their time coding. The automatable, grungier slice is what you should be delegating. The ambiguous or hard work is what you keep driving.

They will not skip the social contract overnight. A human writes code, another human reviews it; the product plugs into that. You ask for code, you decide if it is good, you push it to the team, they review. Over time the surface expands into planning, design, reacting to what is happening in the app, and easier review. Hanson already sees multiple agents collaborating: Codex writes, an Operator agent tests. Non-engineers are already in the mix. While drafting the launch blog with a colleague, Lindsay, a customer wanted to say engineering loved it and it was a power tool for PMs. Alex, who is on the product team and uses Codex so he does not have to bug engineers, still hesitated. The target audience is professional software engineers, not vibe coders. That line did not make the post. Over time, he expects more people to contribute to codebases anyway.

Asked whether the number of professional software developers goes up or down, Alex’s opinion is way up, and he means professionals, not vibe coders. Easier software means more software. Pull up a phone: almost every app was built by a large team for millions of users. Almost nothing is bespoke for you or your team. As that becomes practical, demand for software goes up. Hanson sees a multiplicative factor, not a replacement. Internal power users land 10-plus PRs a day. He cannot picture a world that lowers the bar this much and then needs less software. He also will not pretend certainty. It is an important question. The company pays a lot of attention. They do not know.

Long rollouts add a product problem that autocomplete never had: user intent. Predicting the next microsecond in an IDE is hard; specifying 30 minutes of work is harder, because the user may not know what they want yet. The team still debates the right task granularity, from one-line fixes through refactors you can describe to features you cannot. The practical tip is to use ask mode: let Codex propose a high-level plan, iterate, then send it off for an hour, instead of spending 10 or 20 minutes writing the spec up front. A host’s summary: it really is like working with an intern. On the model side, patience on long rollouts is much better and still finite. Sometimes it comes back after 30 minutes and says, sorry, this is too much, I do not have enough time. That is one of the things it actually says. Also like an intern. Codex is among OpenAI’s longest-running agents; Deep Research held the previous mark.

Alex’s airplane story is the market in one scene. Recently, no Wi-Fi, he thought he might write some code, then decided it was not worth trying. Years ago the genesis of his previous startup was writing code on a plane without Wi-Fi. He would not do that anymore. He thinks an equivalent shift is coming in about two years: coding will look completely different. Today the tools people value most still pair, close in the development environment. The shift, if they can figure out the mechanics, is that the majority of code will be written by agents in their own environments, not only when you think of a task, but connected into the tools you already use. Already at OpenAI, more code is merged by agents, and even more is generated because people kick off a task four times and pick a favorite. How to manage all that code is not 100% clear.

What launched is a research preview, a useful thought experiment, still early. They are proudest of the model and the start of a compute-environment foundation. The ChatGPT UI is a to-do list of questions and code tasks you can go merge. It leans hard into the asynchronous agent. What they want is that you stop thinking about pairing versus delegating. The teammate is ubiquitous: terminal, IDE, issue tracker, the tool that shows errors, maybe it already looked before you arrived and has an opinion. Short question or long, it decides how long to think. Unique to OpenAI, Alex says, they are the makers of ChatGPT. He does not see a day spent choosing among a coding agent, a shopping agent, a taxi agent. One assistant you can ask anything. That is ChatGPT. Power users still get bespoke buttons and lists inside the tools they live in. IDEs do not die. They evolve, from writing code toward landing, reviewing, validating, planning bigger arcs. Internal ritual already looks like: coffee, kick off a few tasks, breakfast, come back to generated PRs, last-mile fine-tune in the IDE on your own vibes.

Sonya, playing the annoying VC, names Claude Code and Jules and asks whether the market converges, and what OpenAI wins on. Alex expects a bit of everything: tools on your computer and tools on their own, with most work in agent compute, while they still invest in accelerating humans at their desks. Hanson’s wedge is context. The hard part of software engineering is encoding the world into requirements and design docs; physical coding is a smaller slice of the lifecycle. ChatGPT has memory, connectors, Operator, Deep Research. Codex gets better at the coding part once it can use that knowledge. Hiring a teammate who will only take a task and emit a PR, then refuses to book a room or lead a brainstorm, would be infuriating. Operator has a browser, Deep Research another flavor of browser, Codex a terminal: roughly a human teammate’s toolbox. The pattern is invest in a specific audience with evals, as they did for developers with Codex and GPT-4.1, then generalize into something everyone can use.

There are also boring, high-leverage things you can do to a codebase. Typed languages. Smaller, better-tested modules. They joke about Alex’s old repo having no tests; they would write it differently today. The internal code name for the Codex project is WHAM, chosen after grepping the codebase, so the agent can search and land on the right files. Call it code, Codex, or agent and search becomes a mess. In the code they still say WHAM a lot, on purpose. Good tests and docs help humans too; now they also help the agent, which is extra incentive to write them. The half-joke future UI, if you are a founder with a couple of co-founders and many agents, looks like TikTok: a vertical feed of agent videos, a customer request, swipe right to fix, left to skip, press and hold for “yes, but italic.” Agents subscribed to company information, proposing, doing, reporting; you curate. Sonya says it is a half joke that finally gave her a picture of reviewing agent work.

Lightning round, compressed. Alex recommends Iain Banks’s Culture novels, unusually positive about a future with AGI. Hanson recommends anything by Richard Sutton; the in-house joke is they read “The Bitter Lesson” every day, which is why Codex gets a terminal and POSIX tools, the least clever way to use a computer. Favorite AI apps: Hanson starts with ChatGPT, then Copilot for daily value. Alex likes AI you do not notice, Linear turning a Slack thread into a bug without saying “AI,” then lands on Waymo. Robotics: both bullish. Breakout category in 2025 besides coding: agents, same answer they heard from Issa and Josh on an earlier show. The shape they are excited about is not a separately provisioned coding agent forever. It is a reasoning model given the tools of the trade, trained to finish a job, eventually one agent with a browser and a terminal that can do several things without you labeling it your coding agent.