What's this about?
A Latent Space episode with Akshay Nathan of OpenAI on how Codex went from a coding tool to powering ChatGPT Work, reaching 10 million combined users within two weeks of the July 9 launch — and why the same agent "harness" now serves both software engineers and ordinary knowledge workers.
What exactly is an "agent harness"?
The "harness" is the scaffolding around a model that makes it an agent: how it plans, runs tools, keeps a persistent computer/environment, spawns sub-agents, and produces artifacts (documents, sites, spreadsheets). OpenAI's insight was that the same harness powering Codex for developers also works for non-developers doing knowledge work — so ChatGPT Work and Codex share it, differing mainly in UX and defaults (like how much Git visibility or sandboxing you see).
Why is this worth noting?
The framing — "100× more people use code than can write it" — captures why labs are racing to turn coding agents into general work agents. Notably, the episode also flags that this same sandbox tech was what an unreleased OpenAI model broke in the Hugging Face incident, tying the launch story back to the day's security thread.