Anthropic’s Claude can coordinate up to 1,000 agents

What Anthropic tested

The company planted 70 bugs in a 116,000-line codebase. A single agent found 14 to 27 bugs per run; a dynamic workflow consistently found 66.

That is a sizable gap, but it comes from Anthropic’s own test, and the article does not specify how the bugs were distributed or how the results were measured. Anthropic says performance on other kinds of tasks remains unknown and recommends testing the system against users’ own workloads.

The open question is cost

Managed agents have been available for some time; dynamic workflows are the new addition. Their potential efficiency is less clear. An OpenAI senior engineer recently called agent swarms a huge waste of tokens, and Anthropic itself warns that these workflows can consume many tokens. It recommends starting with small tasks.

I think the bug-finding result makes a case for testing parallel agents on work where missed errors are costly. It does not yet show that coordinating up to 1,000 agents is practical: the announcement gives no token-cost comparison for the test, and no results from other task types. The useful number for teams deciding whether to adopt this is not the agent limit, but how much extra detection costs.

To enable dynamic workflows, select agent type multiagent_20261001. Setup is available in Anthropic’s documentation or with /claude-api managed-agents-onboard in Claude Code.