Glossary · term

Claude Cowork

Claude Cowork is Anthropic's task-delegation surface for non-coding knowledge work. A user gives it a goal and grants selected files, connectors, browser or application access; Cowork plans and executes multi-step work, can coordinate parallel workstreams, and returns artifacts for review. Desktop is generally available, while some cloud, web, mobile and computer-use capabilities remain beta or research preview.

Products2026-01-12Wave 3 · 2025–26Maturity: 3/5

Origin and context

Cowork launched on 12 January 2026 as a local, isolated-VM preview for Max subscribers on macOS. It expanded through plugins, scheduled tasks and computer use before desktop general availability on macOS and Windows on 9 April. Anthropic later added remote cloud sessions and beta web and mobile access. These milestones matter because the execution location, supported surface and approval controls changed after launch.

Sources: s1, s2, s3

Why it matters

Cowork packages an agent loop for people who want completed documents, analysis, file transformations or browser work without operating a coding terminal. It makes the delegation boundary visible through a plan, progress, artifacts and steering. Independent early tests found file organization, conversion, reporting and browser tasks useful, while also showing why access scope and review remain part of the product rather than optional deployment details.

Sources: s2, s5, s6

Example

A user connects one project folder and a read-only research connector, asks Cowork to compare source documents and produce a cited brief, reviews its plan, and watches for unexpected file or website access. The user checks the generated document before sharing it. For scheduled or write-capable work, the team narrows permissions, chooses an approval mode deliberately, logs activity and avoids irreversible or high-consequence actions without human review.

Sources: s2, s4

How it differs

Computer Use

Computer use is a capability for operating graphical interfaces. Cowork is the broader product surface and prefers connectors or browser integrations before using screen control, which remains a research-preview option.

Claude Managed Agents

Claude Managed Agents is a developer platform for running API-created agents. Cowork is an end-user and enterprise knowledge-work product with its own interface, permissions, sessions and artifacts.

ChatGPT Workspace Agents

Workspace Agents is a separate OpenAI product identity. Similar task-delegation patterns do not make the two products aliases or prove that either one caused the other.

Cloud Agents

Cloud agents are a deployment category. Current Cowork sessions may run in Anthropic's cloud, while legacy local desktop sessions use an on-device loop plus an isolated VM; the product is not defined solely by hosting location.

Maturity and evidence

Maturity is 3. Cowork has a dated product history, desktop general availability, paid-plan distribution, extensive operating and enterprise documentation, multiple supported surfaces, observability controls and independent hands-on use. It is not rated higher because cloud delivery and computer use still carry beta labels, behavior changes quickly and no agent can eliminate action or prompt-injection risk.

Sources: s1, s2, s3, s4, s5

Limits and open questions

Cowork's output quality depends on the model, instructions, available tools, source material and task framing. Isolation constrains where code runs; it does not prevent a model from misusing a file, browser or connector the user authorized. Malicious content can still attempt prompt injection, and Skip mode removes automatic action screening. Cloud sessions process opened local files on Anthropic's servers, while local sessions have a different architecture. Product availability, plan limits and interfaces can change. Consequential messages, purchases, legal or financial decisions, sensitive records and destructive edits require separate domain controls and human verification.

Sources: s2, s3, s4, s5, s6

Related terms

References

Last updated: 2026-09-07

In the Skills Atlas

This term is also covered in the Skills Atlas as computer use ai skill.

In the Skills Atlas

This term is also covered in the Skills Atlas as agentic planning task decomposition skill.

In the Skills Atlas

This term is also covered in the Skills Atlas as multi agent coordination patterns skill.

In the Skills Atlas

This term is also covered in the Skills Atlas as agent sandboxing skill.

In the Skills Atlas

This term is also covered in the Skills Atlas as document ai skill.