Skip to main content

Cowork & Agent Teams

Advanced
What you'll learn
  • What Cowork is — an agentic desktop workspace where you direct the goals and the loop is provided for you
  • What Agent Teams is — multiple agents dividing one goal, each with a role and tools
  • How both relate to the three things you can build yourself: subagents, the Agent SDK, and Managed Agents
  • Why more autonomy means more supervision, not less

Beyond a single agent, Anthropic has been shipping product-level surfaces for letting agents do sustained, collaborative work: Cowork (an agentic desktop workspace) and Agent Teams (multiple agents collaborating). This page is a high-level map — verify specifics against the official docs, since these evolve quickly.

Claude Cowork​

Think of it as a workspace where an agent does real, multi-step work alongside you — operating over files and tools on a longer horizon than a single chat turn, with you supervising. It's the consumer/pro-facing cousin of building an agent on the API: the loop is provided, you direct the goals.

Agent Teams​

Where one agent isn't enough, multiple agents collaborate — dividing a goal, each with a role and tools, coordinating toward a result. Conceptually it mirrors Claude Code's subagents, but as a product surface for sustained, multi-agent collaboration rather than a single delegated subtask.

How this relates to the rest of the site​

Four ways agents get work done
Press Enter or Space to flip the card. Use the left and right arrow keys to move between cards.Term shown.
1 / 5

Quiz​

Check yourself

0/3
  1. You want an agent loop you don't have to run or schedule yourself, called from your own backend. Which option does this page point you to?
  2. What is the relationship between Agent Teams and Claude Code's subagents, as described here?
  3. Why does this page treat supervision as the constant across both surfaces?

The constant: supervision​

More autonomy, more care

Multi-agent, long-horizon work amplifies both value and risk. Keep humans in the loop on consequential actions, scope tool access tightly, and verify outputs — see Responsible Use and Securing Agents.

Key takeaways
  • Cowork and Agent Teams are product surfaces: the agent loop is provided, and your job is directing goals and supervising.
  • They are the product-level cousins of what you can build yourself — subagents for in-session delegation, the Agent SDK for full control, Managed Agents for a hosted loop.
  • These surfaces move fast. Treat this page as a map and confirm specifics against the official docs.
  • Supervision is the constant: human approval on consequential actions, tight tool scopes, verified outputs.

Next​