Claude Managed Agents dynamic workflows: 'completed' isn't 'passed'

Anthropic just put a 1,000-agent fan-out behind one API flag, and the most important line in the docs is about failure. On Oct. 9, Anthropic shipped dynamic workflows in Claude Managed Agents as a beta. The agent writes a small program that runs many agents in phases, the server runs it in the background, and you get one combined result. The Workflow runs docs say a run can end completed even when work on its threads failed. That status means the program finished. It does not mean the work passed.

What actually happened

Per the Claude Platform release notes and the multiagent orchestration docs:

• How to turn it on: set the agent's multiagent field to type multiagent_20261001 with workflows enabled, under the managed-agents-2026-04-01 beta header. You tell the agent in its system prompt when to start a run.
• What a run is: a program the agent writes for the job, such as a review of hundreds of documents. Each agent works in its own session thread, which you can list, read and stream through workflow_run.* events.
• Limits: up to 1,000 agents over a run's whole life, 64 threads working at once (not guaranteed), a 24-hour default lifetime, and 10 open runs per session by default.
• Cost: no separate price for a run. Every agent's tokens bill at normal model rates, and Managed Agents pricing adds $0.08 per session-hour of runtime.
• Budget: you can only set a session budget when you create the session. At the budget, open runs pause. Each thread finishes its current model request first, so a run can overshoot by one request per working thread.

Why it matters for your business

The pitch is easy: point an agent at 400 vendor contracts, a product catalog or a codebase and let it split the work. That is a real capability. Three details in the docs decide whether it saves you money or burns it.

"Completed" is a program status, not a QA status. If your code reads the run result and moves on, it can ship a report where 30 of 400 documents were never reviewed. Read each thread's events, count failures, and fail the job yourself when the count is above zero.

Set the budget at creation, every time. You can't add one later. A fan-out with no cap is an open tab on your API key.

Make your tools safe to call twice. The docs say a run can create more than one thread for the same piece of work and can retry a failed agent on a new thread. If an agent's tool sends an email, writes to Shopify or posts a refund, a duplicate call is a duplicate action. Use idempotency keys, or keep write tools out of the workflow and let a human or one final step apply the changes.

Run work also counts against your normal Messages API rate limits, so a big run can slow your customer-facing features on the same key. Use a separate workspace or key for batch work.

Sources: Claude Platform release notes, Workflow runs docs, Multiagent orchestration docs, Claude pricing.