OpenAI’s ChatGPT Work, released July 9, is not one product wearing a new name. It stitches together at least seven things that used to ship separately: ChatGPT’s conversational mode, the Codex app, the Codex harness, the original Codex cloud agent, ChatGPT agent, the Atlas browser and OpenClaw. That is less a launch than a consolidation, and it says something specific about the last few years at OpenAI: its footprint of agent products had outrun the connective tissue holding them together, and Work is the seam the company is now sewing shut.
The account below comes from Shlok Khemani’s reporting for Latent Space, an outside reconstruction built from days spent poking around inside Work rather than a briefing supplied by OpenAI. OpenAI has not documented how these pieces fit together publicly, so most of what is known about Work’s internals had to be reverse engineered by an analyst, not handed down by the company.
ChatGPT Work sits on top of the Codex harness. That means it shares Codex’s models, its sub-agent architecture, its ability to control a browser and its willingness to stay on a single task for hours at a stretch. The interface, though, strips out the git controls and diff traces that would otherwise mark it as a coding tool to someone who has never opened a terminal. Each task runs inside an isolated cloud microVM. Pro accounts get eight CPUs, 20 gigabytes of RAM and a 64 gigabyte disk. Plus accounts get 14 gigabytes of RAM. Every task keeps a working directory where the agent can create folders, install dependencies and write files with the freedom of an ordinary computer.
Continuity between tasks does not run through that computer, though. It runs through a separate ChatGPT product layer. A tool called Personal Context queries prior conversations on request, and files pass through a Library that keeps its own master copy. Khemani surfaced a specific failure mode here: an uploaded file lives in two places at once, a working copy inside its original thread and the version of record stored in the Library, and the two do not stay synchronized. If the Library file changes afterward, reopening that original thread surfaces the outdated version it saved at the start, not the newer one.
Khemani offers his own read on why OpenAI built it this way, and he flags it as inference rather than confirmed design intent. Rebuilding Conversations, Library, Personal Context and Memory from scratch inside the computer would mean re-architecting a stack already serving hundreds of millions of people. Keeping the computer separate from those systems also works as a guardrail against the kind of unrestricted, single-environment access that made OpenClaw powerful and risky at once.
The browser shows the same pattern. Work controls a separately hosted Chrome instance through tool calls rather than a browser on the agent’s own machine, with a persistent profile that carries logins and preferences across tasks and a permission ledger tracking which sites it may touch. The limits are concrete. Amazon’s US site turned it away, flagging the agent as an unsupported “session or client.” Separately, a Google Photos album copy timed out repeatedly. Both jobs completed once Khemani ran them in Work’s local desktop mode instead.
Discovery is the weakest layer. OpenAI’s plugin system, its latest attempt at a connector model since 2023’s Plugins, now lists more than 1,000 entries in the Plugin Directory. Yet Work never suggested an installed travel plugin when Khemani searched for flights and hotels, and naming Expedia directly did not surface it either.
The forward-looking claim carries real weight. OpenAI’s Greg Brockman has said Chat and Work are set to combine before the year is out, which means the interface being tested on knowledge workers now becomes the default for a chatbot that reportedly reached a billion monthly users in June. What changes is the nature of the default interaction. Today a ChatGPT session is a conversation. Once merged, the default session becomes an agent that can act: book things, edit files and drive a browser inside a permission system most users have never seen. Khemani’s own speculation about what follows, a coordinating layer above individual tasks and full autonomous execution without a human clicking to start it, is roadmap guessing rather than shipped behavior.
Work is a genuine consolidation of years of scattered OpenAI experiments, and it reads as more coherent than the sum of its parts. But rolling it out to a billion Chat users before settling whether the cloud computer becomes a user’s primary machine, and how much sovereignty the agent gets over it, means OpenAI will be making that call live, in production, in front of its entire audience.
Shlok Khemani reported this reconstruction of ChatGPT Work for Latent Space on 4 August 2026.
