Claude Managed Agents Now Fan Out to a Thousand Agents

Anthropic's developer account introduced dynamic workflows for Claude Managed Agents on October 9, and the pitch is blunt: one agent writes a program that runs up to a thousand other agents, in phases, and merges what they return.

What changed

Until now, Managed Agents could hand work to other agents in one way: delegation. The lead agent called a subagent, read the report, and decided the next step turn by turn, with every result landing in the same context window.

Dynamic workflows move the plan out of the conversation and into code. The lead agent writes a workflow - a program that defines the phases, passes results between agents, loops until a check passes, and handles failures - and Anthropic's servers run it in the background as a single workflow run. The main session thread stays free to talk to the user and report progress.

Developers turn it on per agent. Setting multiagent.type to multiagent_20261001 enables both subagents and workflows by default under the managed-agents-2026-04-01 beta header. The limits are documented and hard: up to 64 agent threads working at once, up to 1,000 agents started over a run's life, a 24-hour default run lifetime, and 10 open runs per session. Runs cannot nest, and only the primary thread can start one.

The test Anthropic published

The number that traveled: Anthropic planted 70 bugs in a 116,000-line codebase. Across three runs, a single agent found 14, 15 and 27 of them. A dynamic workflow found 66 in each of its three runs.

Two things are worth separating there. Coverage: 66 of 70 against a best case of 27. And variance: the single agent nearly doubled its result between runs, while the workflow returned the same count every time. That consistency is what you get when orchestration is a script instead of a judgment call. Anthropic has published one test, on its own codebase, with no independent replication yet.

The bill is the real constraint

A workflow run has no separate price. Every agent in it spends tokens at its model's rates, on top of Managed Agents session runtime that outside reporting puts around eight cents per session-hour. When a session hits its budget, open runs pause.

That is the tradeoff the whole category is arguing about. One OpenAI developer has already called multi-agent swarms "massive token waste." Anthropic's own guidance is to start with a scoped slice, because a workflow that fans out to hundreds of agents is just as happy to burn a month of budget in an afternoon.

There is a second trap in the documentation: a run that ends as completed means the program finished, not that the work passed. Failed agents and unchecked scopes only show up if you read each thread's events. Because sessions are stateful, Managed Agents is also not currently eligible for zero data retention or HIPAA coverage, which rules it out for plenty of regulated work.

Builder impact

The pattern worth copying is not "more agents." It is fan-out with isolation, a program holding the loop, and a hard budget cap on the outside. Each agent in a run works in its own session thread with its own context but shares the session's sandbox and credentials, so idempotent tools and narrow permissions matter more than ever.

Audits, large migrations, cross-checked research and codebase-wide sweeps are the shapes that fit. For teams running automation on n8n, the notable shift is where the orchestration lives: it moves from a canvas you can read into a script the model writes, which makes spend controls and per-thread verification the load-bearing parts of the design. Treat "completed" as a status, never as a result.

Key takeaway: Anthropic's dynamic workflows show that agent orchestration is becoming a program rather than a conversation, and that the quality ceiling of a thousand agents is now limited by your budget guardrails and your willingness to read every thread, not by the model.