Anthropic adds dynamic workflows to Claude Managed Agents, with up to 1,000 parallel agents
Anthropic is adding dynamic workflows to its Claude Managed Agents, bringing multi-agent orchestration to the platform. The managed agent infrastructure has existed for a while; dynamic workflows are the new part. In a dynamic workflow, a lead agent creates a plan, distributes tasks to sub-agents, and merges the results when they are done. Up to 1,000 agents can run in parallel per execution.
The Decoder, which reports the news, calls the cost-effectiveness of the approach debatable. It notes that a senior OpenAI engineer, who is not named, recently called agent swarms a massive waste of tokens. Anthropic answers with its own testing. The team hid 70 bugs in a 116,000-line codebase. A single agent caught between 14 and 27 of them per run, while the dynamic workflow consistently found 66. Whether those gains hold across different task types remains to be seen, so the source advises testing with your own workloads.
To activate dynamic workflows, you select the "multiagent_20261001" agent type. Because these workflows can burn through "a lot of tokens," Anthropic recommends starting small. You can get started through the documentation or by running "/claude-api managed-agents-onboard" in Claude Code.
Key facts
- Dynamic workflows are new in Claude Managed Agents: a lead agent creates a plan, distributes tasks to sub-agents and merges the results.
- Up to 1,000 agents can run in parallel per execution.
- In Anthropic's own test, 70 bugs were hidden in a 116,000-line codebase; a single agent caught 14 to 27 per run, the dynamic workflow consistently found 66.
- Activation is by selecting the "multiagent_20261001" agent type; Anthropic recommends starting small because the workflows can use a lot of tokens.
- The source calls cost-effectiveness debatable and says it remains to be seen whether the gains hold across task types.
Why it matters
Multi-agent orchestration moves from something teams build themselves to a built-in mode of Claude Managed Agents. The pattern is a lead agent that plans, sub-agents that do the work in parallel, and a merge step at the end, scaled to as many as 1,000 agents per execution. The open question is economic as much as technical: the source notes that a senior OpenAI engineer recently called agent swarms a massive waste of tokens, and Anthropic's counter is its own bug-finding test.
Who it affects
Developers and teams that run or plan to run agents on Claude Managed Agents, especially on large codebases or tasks that split into many independent pieces. Anyone who pays for token usage is affected too, since the source says these workflows can burn through a lot of tokens.
How to use it
Select the "multiagent_20261001" agent type to activate dynamic workflows. You can get going through the documentation or by running "/claude-api managed-agents-onboard" in Claude Code. Anthropic recommends starting small, and the source suggests testing with your own workloads rather than assuming the benchmark result carries over. No pricing, token counts or cost figures for dynamic workflows are given.
How solid is it
The performance claim rests on Anthropic's own test, and no independent verification is mentioned. The numbers are specific: 70 hidden bugs in a 116,000-line codebase, with 14 to 27 caught per run by a single agent versus a consistent 66 for the dynamic workflow. The test methodology is not described beyond that: the number of runs, the models used and the single-agent configuration are not stated. The source itself says it remains to be seen whether the gains hold across different task types.
Risks and caveats
Cost is the main one. The source calls cost-effectiveness debatable and quotes Anthropic's own advice to start small because of token use. The test covers a single task type, bug hunting in a seeded codebase, so the result may not generalise. No availability status, rollout date or plan restrictions are stated, and the source does not say whether the 1,000-agent limit applies beyond 'per execution'.
“Whether this is cost-effective seems debatable.”
— The Decoder