Anthropic has added dynamic workflows to Claude Managed Agents, allowing a lead agent to plan a job and distribute parts of it across as many as 1,000 parallel sub-agents. The lead agent then combines their outputs into one result.
The company tested the approach on a 116,000-line codebase containing 70 deliberately hidden bugs. A single agent found between 14 and 27 bugs in each run, while the multi-agent workflow consistently found 66. That result shows the potential benefit of parallel searches on a decomposable task, but it is one vendor test and does not establish similar gains for other kinds of work.
Parallelism also carries a direct cost. Each worker uses model tokens and may duplicate investigation performed by another agent. Anthropic warns that the feature can consume “a lot of tokens” and recommends starting with a small number of workers before increasing the scale.
Developers activate the capability with the `multiagent_20261001` managed-agent type. The infrastructure is intended to handle planning, delegation and result collection rather than forcing an application team to build its own orchestration layer.
The useful comparison is therefore not simply one agent versus 1,000. Teams need to measure whether additional coverage or speed justifies the extra inference cost for their own workload, and whether the lead agent can reliably detect conflicting or low-quality results before merging them.