Unpacking ChatGPT Work: The Agent for a Billion Users

I'll pull the Latent Space article so the paragraphs stick to real names, product details, and numbers.On July 9, 2026, OpenAI launched ChatGPT Work, an agent product for knowledge work that sits beside Chat as a separate mode inside ChatGPT. Within about three weeks, Work and Codex together reportedly crossed 10 million users. Latent Space guest author Shlok Khemani published an external reconstruction of the product on August 4, 2026, mapping how Memory, Proactivity, Scheduling, Browser Use, Plugins, Skills, and Tools fit together. Greg Brockman has said Chat and Work will merge by the end of the year, so Work is effectively a preview of how ChatGPT’s roughly billion-scale user base may use the app next. The HN link that surfaced the piece (item 49187587) had 2 points and 0 comments at capture time.

Under the hood, Work runs on the Codex harness and inherits the same models, sub-agents, and multi-hour task loops, while the UI strips git controls and diff traces that would mark it as a coding agent. Cloud tasks run in an isolated microVM: Pro accounts get 8 CPUs, 20GB of RAM, and a 64GB disk; Plus gets 14GB of RAM. Workspace state is synced to persistent storage and restored onto fresh VMs as needed; each task gets a working directory under /workspace/scratch for files, scripts, and local databases. Continuity across tasks does not come from free roaming of that disk. It goes through ChatGPT product layers: compressed summaries of recent work, a Personal Context tool over chat history, a Library for canonical files, and a synthesized user memory profile injected at task start. Browser control is a separate managed Chrome service driven by tool calls, with a persistent browser profile the agent never sees directly and a permission ledger for which sites it may act on.

For engineers and builders, the important split is sovereignty versus product guardrails. Within a single task the agent behaves like a normal computer user: install packages, write analysis code, produce sheets, docs, slides, and Sites (hosted web apps and dashboards shareable by URL). Across tasks, OpenAI keeps conversations, Library items, Projects, and memory as product-owned abstractions rather than OpenClaw-style always-on full filesystem control. That design reuses the stack that already serves ChatGPT at huge scale and reduces the risk of one agent environment holding every file and transcript. It also means builders integrating or competing with Work must plan for tool-mediated context, stale per-thread file copies that do not sync with Library updates, and opaque scratch directories that the agent will not cross-navigate unless told to.

Market context is consolidation, not a greenfield agent. Work folds years of OpenAI surface area: ChatGPT chat, Codex the app and harness, earlier cloud agents, Operator and ChatGPT agent browser use, Atlas, and the long line from Plugins (2023) through GPTs/Actions, connectors, apps/App Directory, and the March 2026 return of plugins as packages of apps and skills. The July 9 launch turned the App Directory into a Plugin Directory spanning Work and Codex, with more than 1,000 plugins. Desktop Work adds local mode (effectively Codex without code-facing chrome) versus cloud mode that syncs across web and mobile; heartbeats for scheduled re-entry into an existing conversation still work on desktop but were not exposed on web at the time of Khemani’s writeup. Discovery remains weak: Work routes to installed plugins but does not reliably suggest uninstalled ones even when a named service would fit.

Watch the Chat–Work merge for which persistence model becomes default, whether cloud and local machines get real handoff, and whether agents gain more computer-level autonomy or stay boxed by Personal Context, Library, and Memory. Treat scheduled automations (standalone cron-style runs versus heartbeats that reawaken a thread) and proactive task suggestions as the near-term UX shift: today they still need a user click; the next step is completing predicted work unprompted. For plugin and MCP builders, install-time discovery and role-specific skill packs (for example the Sales plugin’s many skills across many apps) matter more than raw tool count until Work starts recommending missing plugins during a task.