Claude: Expert
Cowork is the chat app’s agentic edge: a workspace where you hand Claude a multi-step task and it works across your files, connectors, and the web until the job is done. It left research preview in April 2026 and is included on paid plans. Inside Cowork, computer-use mode goes further: Claude actually controls your machine, clicking, typing, and opening apps. This is the closest the desktop app gets to a real agent loop. Because the actions are real, supervision and verification matter more here than anywhere else. If the agentic loop is the engine, Cowork is where it touches the world.
How Cowork decides what to use
Section titled “How Cowork decides what to use”Computer use follows a deliberate tool priority: it reaches for the cleanest path first and only falls back to raw screen control.
- Connectors (Gmail, Google Drive, Slack): structured and reliable.
- Chrome browser navigation: for the web.
- Direct screen control (click, type, open apps), used only when there’s no connector or tool for the job.
It asks before accessing each app, and investment, trading, and crypto apps are blocked by default. This ordering is also good practice to encourage: prefer a connector or browser path over screen control whenever both exist.
| Good fit for Cowork | Poor fit |
|---|---|
| Scoped, reversible multi-step tasks | Irreversible or financial actions |
| Work a connector or browser can do | Anything that must be exact and unsupervised |
| Tasks you’ll watch and approve | Long unattended runs across many apps |
Run it safely
Section titled “Run it safely”A real agent loop rewards tight scoping. This is verification applied to actions, not just text:
- Scope the task to something low-stakes and reversible. Don’t hand it irreversible or financial actions.
- Approve deliberately. Claude asks before touching each app; keep those prompts on, and don’t grant access to banking, health, or government apps at all. Anthropic’s own guidance is to monitor runs closely because no safeguards are perfect.
- Watch the run. Verification here is reading the on-screen actions and approving per-app and per-write prompts; there’s no test or build gate to catch a mistake for you. Close confidential documents before you start.
- Stop and re-scope if it drifts, rather than letting it improvise.
Working away from your desk
Section titled “Working away from your desk”Dispatch gives you one continuous conversation with Claude across devices: assign a task from your phone, let Claude work on your computer, and pick up the result wherever you are. It’s generally available on Pro and Max in Cowork and Claude Code, and pairs naturally with computer use, as long as your machine stays awake with the desktop app open.
Anthropic is also rolling out Cowork on web and mobile, where remote sessions run on Anthropic’s servers instead of your machine, so work continues after you close the laptop. As of mid-2026 this is in beta, starting with the Max plan; check current availability before you build a habit around it.
What the chat app does not do
Section titled “What the chat app does not do”Cowork is not a headless automation platform. There’s no CI, no headless runs, and no parallel-agent orchestration in the desktop app; that capability lives in Claude Code and the API. If you need unattended pipelines or many agents at once, that’s the tool, not this one. The desktop app’s job is a grounded, supervised workspace; keep your most autonomous work where it can be tested and re-run.
You’re done when…
Section titled “You’re done when…”- ✅ You understand Cowork’s tool priority: connectors, then browser, then screen control
- ✅ You’ve run a scoped, reversible Cowork task with write-approvals on
- ✅ You know Cowork ships on paid plans, but computer use is a Pro/Max research preview
- ✅ You know when to reach for Claude Code or the API instead
Next step
Section titled “Next step”You’ve reached the top of the Claude track. Solidify the thinking behind it, or pick another tool.