Managed Agents dynamic workflows fan out hundreds of tasks
Turn on multiagent_20261001 so an agent can write a background program that runs up to 1,000 agents in phases.
You can let a Claude Managed Agents agent write a workflow that fans work across many agents in the background. Anthropic shipped dynamic workflows in beta on Oct 9, 2026.
What changed#
Before: multiagent meant subagents you delegated to, plus an optional advisor. After: with multiagent.type set to multiagent_20261001, the agent can also write a workflow program. The server runs it in the background while the primary thread keeps talking to you or ends its turn.
Limits: up to 1,000 agents per run, 64 working at once, 24-hour default lifetime, 10 open runs per session. Follow workflow_run.* events on the session stream. Dynamic workflows and subagents are both on by default with this type.
Why it matters#
Reviews, audits, migrations and deep research can fan out instead of sitting in one conversation. Token cost scales with every agent, so set a session budget.
How to use it#
- Create or update the agent with the
managed-agents-2026-04-01beta header. - Set
multiagentto{"type": "multiagent_20261001", "workflows": {"type": "enabled"}}(already the default on this type). - Put a clear system-prompt rule for when to start a run.
- Create the session with a
budgetif you need a hard cap (you cannot add one later). - Send the work as a
user.message. Watchworkflow_run.created, phase events, thenworkflow_run.status_ended.
When you are asked to review more than a few contracts, start a workflow run that reads them in parallel and combines the findings. Review one or two yourself without a run. If one agent fails to read a contract, list it as not covered and do not fail the whole run. Give the run a one-hour time limit.
Good to know#
- Inline agents in a run use the session agent's model. List cheaper agents in
workflows.predefined_agentsif you want mixed models. - A custom tool whose name starts with
ant__blocks every update until you rename or remove it. user.interruptstops the primary turn but does not end runs. Ask the agent to stop them, or wait for the lifetime.- Run events do not fire webhooks. Rebuild state from listed events after a reconnect.
- Watch next: a non-beta graduation, and any change to the 64 / 1,000 / 10 limits.
Sources#
- https://docs.anthropic.com/en/release-notes/overview (Oct 9 entry, checked 2026-10-13 5:56 AM ET)
- https://platform.claude.com/docs/en/managed-agents/workflow-runs (checked 2026-10-13 5:57 AM ET)
- https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration (checked 2026-10-13 5:57 AM ET)
Last checked: 2026-10-13