Claude Managed Agents can now write their own multi-agent workflows, with the server running phases in the background and merging results

A Claude Managed Agents agent can now use dynamic workflows, turning work with many pieces into a program that runs many agents in phases and combines their results. The release notes use reviewing hundreds of documents as the example workload, and say the workflow is written by the agent rather than specified step by step by the caller. The capability is in beta with the managed-agents-2026-04-01 beta header, and it is turned on by setting the agent's multiagent field to {"type": "multiagent_20261001", "workflows": {"type": "enabled"}}.

The server runs the workflow in the background as a workflow run, and each run can be followed through workflow_run.* events on the session's event stream. Because a workflow is a program that runs many agents in phases and combines their results, the multi-agent orchestration is written by the agent itself. To guide the agent, the caller tells it in its system prompt when to start a run, and runs are followed through workflow_run.* events on the session's event stream.

The capability targets agents on Claude Managed Agents, is in beta, and requires the managed-agents-2026-04-01 beta header plus workflows enabled in the multiagent field; the release notes use reviewing hundreds of documents as the example workload.

The release notes give no completion quality on specific tasks and no comparison with hand-built orchestration; workflow runs appear as workflow_run.* events on the session's event stream, and behavioral detail is left to the Workflow runs documentation.