Anthropic adds dynamic workflows to Claude Managed Agents, allowing up to 1,000 parallel sub-agents per execution

Anthropic has added dynamic workflows to Claude Managed Agents, enabling a lead agent to coordinate up to 1,000 sub-agents running in parallel per execution, according to The Decoder, which cites Anthropic.

The Managed Agents infrastructure already existed. Dynamic workflows are the new multi-agent orchestration layer on top of it.

How it works

A lead agent builds a plan, assigns tasks to sub-agents, and merges their outputs once they finish. The ceiling is 1,000 agents running in parallel for a single execution.

To enable it, select the agent type multiagent_20261001. Developers can start through Anthropic's documentation or by running /claude-api managed-agents-onboard in Claude Code.

Anthropic's test results

Anthropic reports the following from its own internal testing:

• The team hid 70 bugs in a 116,000-line codebase.
• A single agent found between 14 and 27 bugs per run.
• The dynamic workflow consistently found 66.

These figures are Anthropic's claims. They come from a single task type, a bug-finding benchmark the company designed, and have not been independently verified. It is unclear whether the gains carry over to other kinds of work.

The cost question

Anthropic acknowledges that these workflows can consume "a lot of tokens" and recommends starting with small runs. The company has not disclosed specific pricing for dynamic workflows or per-run token figures in the material available. Standard model token rates would presumably apply to each sub-agent's usage, but that is not confirmed. Pricing not yet disclosed.

The cost concern is not hypothetical. A senior OpenAI engineer recently described agent swarms as a large waste of tokens. Anthropic's bug-hunting result is a direct counterpoint, though it measures detection rate rather than cost per bug found, which is the figure that determines whether the approach pays off.

Not disclosed

• Which Claude models power the lead agent and sub-agents
• Pricing or any surcharge for dynamic workflows
• Token consumption in the 70-bug test
• Rate limits or concurrency quotas by plan
• Whether the 1,000-agent ceiling is adjustable

What this means

This is a product feature, not a new model. It moves Anthropic from offering agent infrastructure to offering built-in orchestration, which developers previously assembled themselves with frameworks and custom code.

The benchmark result is plausible for a task that splits cleanly. Bug hunting across a large codebase is highly parallelizable, since each agent can inspect a separate region. Tasks with tight dependencies between steps, such as sequential reasoning or refactors touching shared state, are less likely to see similar gains, and merge errors could offset them.

The missing metric is cost per correct result. A run that finds 66 of 70 bugs is only better than a single agent's 14 to 27 if the token bill is justified by the value of the extra findings. Until Anthropic or independent testers publish that number, teams should run small pilots on their own workloads and track tokens against outcomes, as Anthropic itself advises.