AI news · October 12, 2026
Anthropic lets Managed Agents spawn up to 1,000 agents per run
Claude Managed Agents can run up to 1,000 agents per execution, with 16-way concurrency in the cookbook; use the beta on small jobs first because token cost can scale sharply.
- maximum agents per run
- 1,000
- concurrent agents in the cookbook example
- 16
- bugs found in one reported dynamic-workflow test
- 66
Anthropic added dynamic workflows to Claude Managed Agents in beta. Instead of handing the developer a fixed list of subagents, Claude can write a workflow for the job, run work in phases and combine the results in a server-managed background run. The API uses the multiagent_20261001 agent type with workflows enabled; Anthropic's cookbook documents up to 16 concurrent agents and the managed-agent docs set a maximum of 1,000 agents per execution.
AI Tool Herald reports an internal code test with 70 hidden bugs in a 116,000-line codebase: one agent found 14–27 per run, while the dynamic workflow found 66. That is a narrow vendor test, not a general productivity figure. The practical alternative is a hand-built queue, which gives better cost control but more orchestration work. Start with a small sample, cap tokens, record each sub-result and inspect failures before allowing a large fan-out.
What you can do with it
Read the workflow-runs docs and cookbook, enable the beta header, and run a bounded code-review or document-comparison job. Set an agent cap below the 1,000 maximum, store per-agent outputs, and compare total cost and useful findings with one strong agent before scaling.
Our take
This makes parallel agent work easier to ship, but it also makes runaway spend easier to create. The 66-bug result is attractive because it names a real test, yet it is still one codebase and an internal report. The product win is for teams that can measure every branch and discard weak work automatically.
Source: Anthropic / AI Tool Herald — Made With Models writes the brief; the reporting is theirs.