The Decoder reported on 9 October that Anthropic is adding dynamic workflows to Claude Managed Agents. A lead agent makes a plan, hands tasks to helper agents, and merges their results. Up to 1,000 agents can run at the same time in one execution. In Anthropic's own test, the team hid 70 bugs in a codebase of 116,000 lines. A single agent found between 14 and 27 per run. The dynamic workflow consistently reached 66.
Users switch the feature on by choosing the agent type "multiagent_20261001". Anthropic warns that these workflows can use many tokens, the units AI models are billed by, and advises starting small. The Decoder says it is not yet known whether the bug-test gains hold for other kinds of tasks. It also notes that a senior OpenAI engineer recently called agent swarms a big waste of tokens. The report gives no prices, token counts or wider rollout details.
Teams that build or buy AI agent tools now have a way to spread one job across many agents inside Anthropic's managed platform. Whether that pays off depends on token cost against results, and The Decoder suggests testing it on your own workloads.