Claude Managed Agents can now write its own multi-agent workflows, and a completed run does not mean the work passed

On October 9, Anthropic added dynamic workflows to Claude Managed Agents, in beta. For a job with many pieces, such as reviewing hundreds of documents, the agent writes a small program that runs many agents in phases and combines their results, and the server runs that program in the background. Three days later, at AI Engineer New York, two Anthropic staff walked through it in a workshop that Latent Space wrote up: the idea started in Claude Code, and the lead agent now chooses how to organize the other agents itself instead of following a roster a developer set up.

The limits are in the workflow runs documentation: 64 agents working at once in one run, a number the documentation says it does not guarantee, 1,000 agents over a run's life, a 24-hour default lifetime and 10 open runs per session. The more useful line sits further down. A run that ends completed only means the program finished. The documentation says plainly that a run can end completed even though work on some of its threads failed. The work is done when every run has ended and the session then goes idle with end_turn; until then, an idle session proves nothing.

That matters to Jon because he runs a fleet of agents on Claude, and the problem this documentation describes is the one his system was built around. Jon's agents send coding work to separate workers, each in its own copy of the repository, and the coordinating agent reads the actual diff before anything is committed. A worker saying "done" is treated as a claim to check. Managed Agents now draws the same line in its own reference pages.

Money works the same way. A session budget stops new model requests once the session's cost at public list prices reaches the cap, but the request already running still finishes, so the final cost can land a little past it. Under Managed Agents pricing, a session pays model token rates plus $0.08 per hour of running time, and the Batch API discount does not apply.

How to use it

Use the managed-agents-2026-04-01 beta header and set the agent's multiagent field to {"type": "multiagent_20261001", "workflows": {"type": "enabled"}}. Tell the agent in its system prompt when a run is worth starting. Pass a budget when you create the session; "amount": "125" means $1.25, written in whole cents. Then:

• Track every workflow_run.status_ended, then wait for end_turn.
• Read each thread's events for failures before trusting the answer.
• Make every tool safe to call twice, because a failed agent is usually run again on a new thread.
• After a dropped connection, rebuild state from the paginated event history; the stream does not replay what you missed.

The catch: multiagent orchestration has no approval step before a run starts. A prompt asking the agent to describe its phases and wait is the only brake, and nothing enforces it.

What nobody knows yet

Whether flows beat one agent on quality or cost. The workshop presenters said a stock-scanner flow found more details and cost more than a single agent; Latent Space reports no numbers for either. Nobody outside Anthropic has published a measured comparison.

What to watch

When the feature leaves beta, the limits may change; the 64-agent figure is already marked as not guaranteed. Watch the release notes for that, and for whether Claude Code subscriptions get the same controls.