Anthropic has released dynamic workflows for its Claude Managed Agents, enabling a lead AI agent to create plans, distribute tasks, and merge results from up to 1,000 parallel sub-agents. The feature can be activated using the “multiagent_20261001” agent type via the Claude API or Claude Code onboarding.

While critics like OpenAI engineers argue multi-agent swarms waste tokens, Anthropic reported internal benchmark gains. In a test involving 70 hidden bugs across a 116,000-line codebase, a single agent caught 14 to 27 bugs, whereas the dynamic workflow consistently identified 66.

Anthropic warned that these multi-agent workflows consume a high volume of tokens and recommended that developers start with smaller workloads to assess cost-effectiveness for their specific use cases.

Why it matters

  • Enables complex, large-scale code reviews and automation by running up to 1,000 parallel Claude sub-agents.

  • Significantly increases token consumption and compute costs, requiring careful evaluation of ROI versus single-agent workflows.

  • Provides a structured, managed framework for orchestration, reducing the need for custom multi-agent harness development.

Source: the-decoder.com