Claude Managed Agents can now write their own background workflows
Anthropic's Managed Agents API lets an agent write a workflow that runs many other agents in phases and combines their results, instead of delegating one subagent at a time.
Anthropic has added dynamic workflows, in beta, to its Claude Managed Agents API as of 9 October. Until now, an agent running under this API could only delegate to subagents one task at a time. Now it can write a whole workflow itself: a program that runs many agents in phases and combines what they return, which the server executes in the background as a “workflow run” while the main session keeps talking to the user or checks in on progress.
You turn it on by setting an agent’s multiagent.type to multiagent_20261001, under the managed-agents-2026-04-01 beta header, with workflows enabled (it is on by default for that type):
{
"multiagent": {
"type": "multiagent_20261001",
"workflows": { "type": "enabled" }
}
}The agent decides for itself whether a job needs a run, based on how you brief it in the system prompt, and you follow progress through workflow_run.* events on the session’s event stream. Anthropic points to document audits, migrations and cross-checking many sources as good fits.
Why it matters: every agent in a run spends tokens, so Anthropic’s own advice is to set a session budget before you turn this on, since runs can’t get one added after the fact. If you’re building on this API, this changes how much orchestration logic you need to hand-write yourself.