Claude can run up to 1,000 AI agents on a single task

What's new

Anthropic has added dynamic workflows to Claude Managed Agents. This is Anthropic's platform for AI agents, meaning programs that can solve tasks on their own in multiple steps. The principle is that a lead agent makes a plan, hands out the work to a group of subagents, and finally gathers the results. Up to 1,000 agents can be used on a single job.

The feature is live now. Developers choose the agent type "multiagent_20261001" in Anthropic's documentation, or they run the command "/claude-api managed-agents-onboard" in Claude Code. According to other outlets, a similar feature was opened as a research preview in Claude Code in late May 2026. There, Claude writes a small script itself that controls the agents. According to those reports, no more than 16 agents run at a time, while a single workflow can reach 1,000 in total.

What's smart about it

Think of it as a project manager who doesn't do all the work personally, but splits it among many employees who work at the same time. Large tasks become both faster and more thorough, because each agent can focus on a small piece.

Anthropic tested it on a codebase of 116,000 lines in which 70 bugs had been hidden. A single agent found between 14 and 27 bugs per attempt. The dynamic workflow consistently hit 66 of the 70. Anthropic itself stresses that the result may look different for other types of tasks. It is also an internal test, not an independent measurement.

Cheaper or better?

It appears to be better when the task is large and can be split up. Cheaper it is hardly. Anthropic itself warns that workflows can burn through many tokens (tokens are the chunks of text the AI model reads and writes, and which you pay for). The company recommends starting small.

A concrete price for the feature is not stated in the material I have been able to find. It is also not stated whether it requires a particular subscription, or when it will reach regular users in Denmark and the EU. You will probably pay for the tokens all the agents use, so the bill grows with the number of agents.

Against the competition, the picture is mixed. The Decoder notes that a senior engineer at OpenAI has, according to the source, called agent swarms "a massive waste of tokens." Anthropic is thus betting on a method that not everyone agrees with.

What it's good for

The feature is first and foremost for developers and IT departments. It suits tasks such as reviewing a large codebase for bugs, big cleanups or rewrites of code, and other jobs that can be cut into many independent pieces. One example from the reports is a developer who, according to one source, had around 750,000 lines of code written in eleven days. The figure comes from a third party and has not been verified by us.

For small, everyday tasks such as writing an email or a summary, 1,000 agents make no sense. A single agent is both enough and cheaper.