Unpacking ChatGPT Work: the Agent for a Billion Users

OpenAI didn't so much launch a product on July 9 as detonate one. Three new models spread across fourteen configurations, the ChatGPT and Codex desktop apps welded together, and cloud agents pushed out to anyone with a subscription rather than just developers who know what a git diff is. Three weeks later, Work and Codex combined have reportedly pulled in 10 million users, and Greg Brockman has already confirmed that Work and regular chat mode will merge by the end of the year. That's the real headline here: this messy, half-finished thing is a preview of what a billion weekly ChatGPT users will eventually be handed.

Strip away the branding and Work is Codex with the incriminating evidence removed. Same harness, same models, same ability to grind on a task for hours, same sub-agents doing browser work behind the scenes. What's gone is the UI chrome that would tip off a non-coder that they're driving a coding agent — no git controls, no diff traces. Every task spins up inside an isolated microVM, and OpenAI isn't stingy about the hardware: Pro accounts get 8 CPUs, 20GB of RAM and a 64GB disk, Plus users get 14GB of RAM. That's a genuine computer doing genuine work in the background, not a chatbot pretending to think.

Where it gets interesting is memory. OpenClaw-style agents get a computer of their own that persists across every conversation, building up state like a real employee would. Work does something more cautious. Each task gets a scratch directory where the agent can install software, write scripts and build files freely, but that directory doesn't automatically talk to other tasks. Context moves between them through a separate product layer instead — a compressed summary of recent work, plus a tool called Personal Context that queries chat history on demand. Files live in two places at once, a working copy in the thread and a canonical version in ChatGPT's Library, and the two don't sync automatically, so a thread can end up working from a stale copy without anyone noticing. It's a deliberate tradeoff: less raw agent autonomy in exchange for OpenAI keeping control over sharing, versioning and safety at billion-user scale.

The more forward-looking piece is proactivity. Open a new Work task and it surfaces suggestions pulled from your calendar, email and memory — reasoning through what you probably need before you ask for it. One offered to prep a meeting brief for a call already sitting on the writer's calendar, pulling from Gmail and Calendar without being told to. Scheduled Tasks push this further, letting Work run recurring jobs — daily briefings, weekly recaps, monitoring loops — on a heartbeat that keeps context alive between runs. None of it executes without a human clicking go yet, but the gap between suggesting useful work and simply doing it looks small.

Browser use is where the seams show. The agent doesn't control a browser on your machine, it drives a separately hosted Chrome instance through tool calls, complete with its own persistent, logged-in profile that survives across tasks. That works fine for most sites. It falls apart when the target notices it's talking to a datacenter: Amazon US flagged it outright as an unsupported client, and Google Photos timed out repeatedly on a task that worked instantly in local mode. Work can attempt a CAPTCHA, but that's about where its patience with the open web runs out.