Alex from OpenAI's Codex team: 60-hour autonomous coding sessions, enterprise pull, and the human teammate analogy

Summary

Alex from OpenAI's Codex team used a Microsoft Build appearance to announce two integrations: Codex is coming natively to GitHub, and is now available to Copilot Pro Plus subscribers inside VS Code — no separate ChatGPT account required. The Copilot Pro Plus deal is the more structurally interesting piece. OpenAI is effectively licensing its model and the underlying "harness" — the tooling, run loop, and context management layer — directly through GitHub's subscription, removing the friction of a second OpenAI account for the large installed base of enterprise Copilot users.

The framing Alex returns to repeatedly is the human teammate analogy: Codex is present wherever a developer already works — Slack, GitHub, the CLI, VS Code — the same agent surfaced across all those contexts. The ambition is less about replacing the individual coding session and more about covering the full collaboration surface, from reviewing a PR in GitHub to pushing a change from your phone.

60-hour autonomous sessions

The headline capability Alex shared at the keynote: the Codex team ran the model autonomously for over 60 hours on a single hard task. That's the context window and the harness working together to sustain productive work well beyond a single session length. Alex is careful to frame session length as a lagging indicator rather than a target metric — the actual driver is model intelligence, and longer autonomous runs are a byproduct of smarter models taking on harder tasks.

The 60-hour figure raises an obvious UX question: how do you avoid a developer returning to find 60 hours of wasted work? Alex points to two things the team is actively building toward — steerability, so developers can tune task length and scope, and a check-in workflow closer to the teammate model, where Codex pings the developer for a quick review before continuing rather than running blind to completion.

Enterprise pull

Alex's read on the current user base is that Codex skews toward professional engineers, though a meaningful share of users come in with limited coding experience. The GitHub integration is partly a distribution bet: there are developers living almost entirely inside GitHub's collaboration layer, and making Codex native there meets them where the work already happens. The college CS student who never learns to code outside an AI-assisted environment is the downstream version of that — someone who will spend most of their time commanding fleets of agents, with GitHub as the collaboration layer on top.