Claude now writes its own programs to run up to 1,000 agents, and starts them itself
Claude Managed Agents can now write a program that runs up to 1,000 agents, and the model decides for itself when to set one going. Anthropic's platform release notes put the feature, dynamic workflows, in beta on October 9, behind the managed-agents-2026-04-01 beta header, and describe a workflow as "a program that runs many agents in phases and combines their results". The server runs it in the background as a workflow run.
The Claude wordmark and starburst logo on a beige background. Image: Anthropic
Delegation to subagents came first, one level deep, with at most 25 child threads at a time. A workflow is a different shape: Anthropic's orchestration guide says that with dynamic workflows "Claude writes a program to orchestrate agents without Claude's direct involvement", passing results from one agent to the next in code while the main thread carries on talking to the user.
The agent decides when a run starts
Nothing in the flow asks a developer to approve the launch. "You describe the work in a user.message," the guide says, and "From that, the agent determines whether and when to start a run, so there's no additional API call." The lever on offer is persuasion: you describe when to use a run in that message or the agent's system prompt.
Anthropic is equally plain about what the usual safeguard covers. "Permission policies apply to the tools that a run's agents call, not to starting the run." The tools a fleet reaches for stay governed; raising the fleet does not sit behind a policy. Dynamic workflows can be switched off through the workflows setting in an agent's multiagent block, but with the multiagent_20261001 type both workflows and subagents are enabled by default.
A budget is the brake, and it can be passed
Money is where the limit lives, and it is approximate. "A run has no price of its own," the workflow runs documentation says. "The tokens its agents use are billed like the session's other tokens, at each model's rates." When a session hits its budget, "Every open run pauses, and the session reports idle with budget_reached", and then comes the caveat: "Each thread finishes the model request it already started, so a run can pass the budget by one request for each working thread."
With 64 threads able to work at once, the overshoot is bounded by the thread count rather than by any figure Anthropic publishes. Raising or removing the budget resumes the runs it paused. Working out what one run cost means listing the session's threads and totalling the token counts carrying its identifier, because no endpoint lists runs.
The published ceilings, and the one that is not a promise
Four limits are documented. A run has 64 threads working at once, which Anthropic qualifies: "The API doesn't guarantee this number, so it can change." A workflow starts 1,000 agents over a run's whole life, after which the run ends with thread_limit_error, though "The server can run a failed agent again on a new thread, so a run might have more than 1,000 threads." A run lives 24 hours by default, or for a lifetime the agent sets, and "No event says what lifetime the agent set." A session holds 10 open runs by default, idle ones included.
Beyond those, the guidance turns vague by design: "The server has other limits on workflows, and rules for them, that aren't listed here." Archiving or deleting a session with a run still open "Might return 400" with an error code of workflow_run_open, and "Might also succeed".
An inline agent that a workflow defines uses the model of the agent running the session, so mixing in a cheaper model means creating those agents first and listing them in workflows.predefined_agents. All three pages were read at 21:35 UTC on October 10. Managed Agents spending can come out of the monthly API credits that Max and Team plans now include, and a run's tokens are billed at each model's own rates, starting from Claude Opus 5.5 at $4 per million input tokens.
