Claude Managed Agents launches “Dynamic Workflows”: one agent orchestrates thousands of agents.

Anthropic’s Claude Managed Agents has opened a public beta for “dynamic workflows”: the main agent first writes a plan, then hands out work in stages to a large group of agents running in parallel, and finally aggregates the results. A single run can dispatch up to 1000 agents.

Claude Managed Agents is Anthropic's managed agent service for developers. You don't have to build your own agent loop or sandbox: configure the model, tools, and prompt, and the agent runs in Anthropic's cloud environment. It's a good fit for long tasks that run anywhere from a few minutes to a few hours.

It already supported “sub-agents”: the main agent hands out tasks itself, then reads each sub-agent’s report one by one, with up to 25 running at once in a single session. Dynamic workflows work differently: the main agent writes a program, and the server executes it in stages in the background, passing context and results between agents through the program. This frees up the main agent to keep talking to the user and report progress at any time.

Anthropic shared a set of test data: in a codebase of 116000 lines, they planted 70 bugs. A single agent ran 3 times and found 14, 15, and 27 bugs respectively — the results varied widely. Running it as a workflow 3 times, it found 66 bugs every time.

It works best on tasks that can be split into many chunks: auditing large codebases, migrating code in bulk, reviewing hundreds of contracts, or cross-checking multiple sources during research. The official docs use contract review as an example—one or two contracts the agent can read on its own, but once there are more, you spin up a workflow to read them in parallel.

To use it, set the multiagent type to multiagent_20261001 in the agent configuration (workflows are on by default for this type), then tell Claude to “run a workflow” — it handles both planning and scheduling. In Claude Code, type /claude-api managed-agents-onboard bug-hunter to get started right away, and there are templates on http://platform.claude.com/agent-quickstart too.

Watch your costs. Every agent in the workflow burns tokens, and when hundreds or thousands run at once, the bill climbs fast. Anthropic recommends starting with a small-scoped task, then gradually increasing complexity. The docs also suggest setting a "session budget" when you create a session: once you hit the limit, the workflow pauses automatically and resumes after you raise it. This budget can only be set at session creation time—you can't add it afterward.

Agents defined inline in a workflow use the same model as the main agent by default. If you want to hand some of the work off to a cheaper model, you have to create those agents separately first and then list them in the config.