Anthropic lets one Claude agent spawn up to 1,000 sub-agents

The latest:

A single Claude agent can now write and launch its own multi-stage program that dispatches as many as 1,000 sub-agents, after Anthropic moved Dynamic Workflows into public beta on the Claude Managed Agents platform on 9 October 2026. The company’s documentation describes the job as a workflow run executed in the background on Anthropic servers. The previous architecture allowed one level of delegation and 25 parallel threads.

Details:

• How it works: According to Anthropic’s documentation, the lead agent writes a workflow that runs several agents in stages and aggregates their results. The platform executes it in the background as a workflow run, with the agent deciding on its own when to start based on the user message or the system prompt.
• No approval gate: Anthropic states that no extra API call and no developer approval step is required to trigger a run. A developer’s only lever is describing the circumstances in which a run should be used, which shifts the launch decision to the model itself.
• The stated limits: Up to 1,000 agents may be launched over the life of one run; exceeding that ends the run with a thread_limit_error. Up to 64 threads run concurrently, a figure Anthropic says is not guaranteed. Default run lifetime is 24 hours, with a default of 10 open runs per session and up to 20 predefined agents per workflow.
• Turning it on: Activation requires the beta header managed-agents-2026-04-01 and setting the multiagent field to type multiagent_20261001 with workflows enabled. Only an agent on the main thread can start a run, so runs cannot be nested, and workflows can be disabled through the multiagent block.
• Budget behaviour: When a session hits its budget, open runs pause and the session reports idle status with budget_reached, per the documentation. Threads finish requests already in flight, so a run may overshoot the budget by one request per working thread. Raising or removing the budget resumes the runs.
• Cost tracking: Anthropic does not provide an endpoint that lists runs, so run cost is calculated by summing the tokens of threads tied to its identifier. There is no separate charge for a run itself.
• Pricing: Agent tokens are billed at each model’s rates, plus 0.08 dollars per session hour of Managed Agents runtime. Built-in agents use the session agent’s model; to run a cheaper model, agents must be created first and listed under workflows.predefined_agents.
• Internal test: Anthropic reported that on a 116,000-line codebase seeded with 70 bugs, a single agent found 14, 15 and 27 bugs across three attempts, while a workflow run found 66 in all three. The company did not publish token consumption or the cost of the test.
• Operational caveats: Once a run starts, the agent cannot send follow-up messages to its threads, and the server archives every thread when the run ends. Archiving or deleting a session with an open run may return a 400 workflow_run_open error. Anthropic recommends starting with small tasks.

Background:

Before this release, Claude Managed Agents permitted delegation only one level deep, capped at 25 concurrent sub-threads. That structure required developers to build their own orchestration layer for anything larger, such as reviewing hundreds of documents or an entire repository.

Between the lines:

The bug-finding gap Anthropic reported — 66 versus a best single-agent score of 27 — is the company’s own internal result, with no published token figures. Set against the documented pricing of per-model token rates plus 0.08 dollars per session hour, and Anthropic’s own advice to start small, the accuracy gain carries an unquantified cost. Removing the developer approval step compounds that: the model, not the developer, decides when a 1,000-agent job begins.

What’s next

Watch for the move out of public beta, any change to the 64-concurrent-thread figure Anthropic says is not guaranteed, and whether the company adds an endpoint for listing workflow runs or publishes token costs for its bug-detection test.