Anthropic is adding dynamic workflows to Claude Managed Agents, enabling a lead agent to divide work among sub-agents and bring their findings together. The system can run up to 1,000 agents in parallel in a single execution, opening a way to tackle tasks through coordinated, concurrent work.
How the workflow divides a task
Anthropic’s managed agent infrastructure has existed for some time; the new feature is the dynamic workflow. A lead agent first develops a plan, then assigns tasks to sub-agents. Once those agents finish, the lead agent merges their results.
This design makes the lead agent responsible for organizing the work as well as consolidating the output. The sub-agents handle separate assigned tasks in parallel, which could help when a problem can be broken into parts. The stated limit is up to 1,000 agents per execution, though that capacity alone does not establish how useful or economical the arrangement will be for any particular job.
Anthropic’s bug-finding test
Anthropic tested the approach on a codebase containing 116,000 lines, with 70 bugs deliberately hidden in it. In the company’s comparison, a single agent found between 14 and 27 bugs per run. The dynamic workflow consistently found 66.
Those results suggest that coordinating agents may help with this specific bug-finding task. They do not show that the same improvement will appear in other kinds of work. Anthropic’s test describes one codebase and one task, so teams considering the feature would need to assess it against their own needs.
There is also a question of cost. The source notes that agent swarms have been criticized as a massive waste of tokens by a senior OpenAI engineer. Anthropic likewise cautions that dynamic workflows can consume “a lot of tokens.” More agents may enable broader parallel work, but the reported results do not by themselves show whether the additional token use is worthwhile.
Start with a small workload
Anthropic recommends beginning with a small workload because of the potential token consumption. That gives teams a chance to see how the workflow behaves on a task they care about before scaling up. Results from their own trials can help determine whether splitting the work improves outcomes enough to justify the resources used.
To activate the feature, select the “multiagent_20261001” agent type. Anthropic says users can get started through its documentation or by running “/claude-api managed-agents-onboard” in Claude Code.
What teams should evaluate
The feature’s potential depends on the task: it needs work that can be divided among sub-agents, and the lead agent must combine their results. Anthropic’s reported bug test offers one example of a measurable gain, but the company says it remains to be seen whether those gains carry over to different task types.
For developers evaluating Claude Managed Agents, a practical first step is to choose a bounded workload and compare the dynamic workflow with a single agent. The useful question is not simply how many agents can run at once, but whether coordination improves the result for that workload while keeping token use manageable.