ChatGPT Work agents: 223 tools, headless Chrome, open-internet exec
A public reference maps OpenAI's agent for the first time: parallel subagents, a persistent filesystem, one-click site deploys, looser sandboxing than rivals.
Copy markdown
223 tools, 44 skills, mapped at last
The reference topping Hacker News is the first full map of OpenAI's agent surface: 44 skills and 223 tools spanning .docx/.xlsx/PDF, browser control, image generation, and data dashboards. You can finally weigh it against Claude Code and Cursor on specifics, not vibes.
Its code sandbox has the open internet
Work's execution environment allows open internet access by default — install any package, clone any repo, hit any API. Claude Code's container defaults to a short PyPI/NPM/GitHub allowlist. More reach, more blast radius: treat every fetched page as hostile before you let an agent run.
A real headless Chrome, not screenshots
Agents drive a full headless Chrome instance — load pages, click, type, fill forms, manage tabs and dialogs, take screenshots — with persistent browser profiles and a per-site permission ledger. Logins prompt you without exposing credentials to the model.
Parallel subagents on a shared disk
Work spawns parallel subagents (Sol, Luna, Terra) over a persistent /workspace scratch volume that outlives a single session — closer to a multi-agent harness than a chat window. The Ultra reasoning tier delegates to them more eagerly.
It can ship a whole stateful site
The Sites plugin deploys full web apps to Cloudflare Workers with D1 and R2 backends, and prompts can be scheduled to run on their own. Weekend build: point it at a spec and have it stand up a small stateful app end to end.
The catch: $20/mo and no SDK
Work needs a paid plan — $20/mo Plus and up; free and $8 Go tiers are out. microVMs scale by tier (8 CPU/20GB on Pro, 14GB on Plus), and there's no public programmatic API yet. It's a product surface to drive by hand, not an SDK to wire into CI.