Claude can now split one job across up to 1,000 agents. Set a budget before you try it

Fan-out is an old automation pattern. Anthropic just made the model write the loop for you.

What happened

Anthropic released dynamic workflows for Claude Managed Agents in public beta on October 9. A lead agent writes a plan, distributes tasks to sub-agents in phases, and merges the results at the end, The Decoder reports.

Key details from Anthropic's documentation, as summarized by Mixed:

• A run can start up to 1,000 agents over its lifetime, with up to 64 working at once. Anthropic says that concurrency is not guaranteed and may change.
• You enable it with the multiagent_20261001 agent type. Claude decides whether and when to start a run.
• A run has no separate price. Its agents' tokens are billed at each model's rates.
• When a session hits its budget, open runs pause. Each working thread finishes its current request, so a run can go over by one request per thread.
• Permission policies apply to the tools a run's agents call, not to starting the run.
• Inline agents use the session agent's model unless you predefine cheaper agents.

Anthropic's own test: 70 bugs hidden in a 116,000-line codebase. A single agent caught 14 to 27 per run, while the dynamic workflow consistently found 66. The company warns workflows can use "a lot of tokens" and suggests starting with a scoped task.

My take

This is the split, process and merge pattern we already build in n8n and Make, except the model writes the plan. It pays off when the work is wide, the pieces are independent and the output is easy to check:

• Auditing thousands of CRM records for duplicates and missing fields
• Reviewing a folder of contracts against one checklist
• Enriching a lead list where each row is its own task

It is a poor fit for sequential work where step 3 depends on step 2.

Before the first run, set three things:

1. A session budget. It is your real spend cap, so set it low and raise it later.
2. Cheaper worker agents. Predefine them so the sub-agents do not all run on your most expensive model.
3. Tool permissions. Since Claude chooses when to start a run, your control lives in what the agents are allowed to call.

The 66 of 70 result comes from one task type in Anthropic's own test. Run it on your own data and compare cost per finished job before you commit.