Anthropic's agents can now write their own orchestration programs, and the published limit is 64 at a time rather than a thousand

What shipped

Anthropic has added dynamic workflows to Claude Managed Agents, documented as a beta behind the managed-agents-2026-04-01 header. The orchestration docs describe it as one of three ways an agent can hand work to other agents: it can delegate to subagents and read what each reports, it can consult an advisor model, or it can write a workflow.

A workflow is a program. The agent writes it for the job you described, and the server runs it in the background as a single workflow run, divided into named phases, with each agent working in its own session thread you can list and stream. The point is where the intermediate results live: in the workflow’s own variables rather than in the agent’s context window, so the conversation holds the final answer and nothing else.

You enable it by setting the agent’s multiagent.type to multiagent_20261001. With that type, delegation and workflows are both on by default.

The numbers, as the documentation states them

Coverage of the launch, including The Decoder’s, led on 1,000 agents running in parallel. Anthropic’s own limits table says something different, and the difference matters if you are sizing a job:

• 64 threads working at once in one run. The docs add that the number is not guaranteed and can change.
• 1,000 agents a workflow starts over the run’s whole life. Past that the server starts no more and the run ends with thread_limit_error.
• 24 hours of run lifetime by default, or a shorter one the agent sets. Time spent waiting on your client counts toward it.
• 10 runs open at once in a session by default, with idle runs counting.
• 20 unique agents in each predefined list, for subagents and workflows separately.

So 1,000 is the budget for a whole run, and 64 is how wide it gets at any moment. Separately, delegation to subagents is one level deep and a session can hold at most 25 child threads at a time.

What it costs

A run has no price of its own. Its agents’ tokens are billed like the session’s other tokens at each model’s rates, which is why the docs tell you to set a session budget when you create the session — you cannot add one later. At the budget every open run pauses and the session reports idle with budget_reached; each working thread finishes the request it had already started, so a run can overshoot by one request per thread.

Only the agent starts a run. No API call does, and no event you send ends one, though archiving the session can.

The result everyone quoted

The Decoder reported an Anthropic test in which 70 bugs were planted in a 116,000-line codebase: a single agent found between 14 and 27 per run, while the dynamic workflow found 66 every time. We could not locate that experiment in the documentation Anthropic has published publicly, so it is reported here as The Decoder’s account of Anthropic’s own material and not as a verified figure. It is a vendor result either way.

Two things called the same thing

Note that Claude Code has a feature with the same name and different limits: a locally executed JavaScript runtime with up to 16 concurrent agents by default, 4,096 items per fan-out call, and the same 1,000-agents-per-run ceiling. The Managed Agents feature is the server-side one.

What to watch

Whether Anthropic publishes a reproducible evaluation for the orchestration itself rather than for an unnamed internal codebase, and whether the 64-thread concurrency figure holds once the beta header comes off. The docs already warn that it may not.