Anthropic Launches Dynamic Workflows in Claude Managed Agents: Orchestrating 1000+ Parallel Agents for Large-Scale Code Audits

Claude Managed Agents Launches "Dynamic Workflows": One Agent Orchestrates Thousands

Anthropic has opened public beta for "dynamic workflows" in Claude Managed Agents: a lead agent first drafts a plan, then distributes tasks in phases to a large fleet of agents running in parallel, and finally aggregates the results. A single run can orchestrate up to 1,000 agents.

Claude Managed Agents is Anthropic's managed agent service for developers. Developers don't need to build their own agent loops or sandboxes; they simply configure models, tools, and prompts, and the agents run in Anthropic's cloud environment. This setup is ideal for long-running tasks lasting from minutes to hours.

The platform previously supported "sub-agents": the main agent assigns tasks and sequentially reads reports from sub-agents, with a maximum of 25 concurrent sub-agents per session. Dynamic workflows take a different approach: the main agent writes a program that the server executes in the background phase by phase, passing context and results between agents via the program. This frees up the main agent to continue conversing with the user and reporting progress at any time.

Anthropic provided test data: In a codebase of 116,000 lines containing 70 bugs, a single agent run three times found 14, 15, and 27 bugs respectively, showing high variance. Using dynamic workflows, three runs each found 66 bugs.

This feature is best suited for work that can be broken into many chunks: auditing large codebases, batch code migrations, reviewing hundreds of contracts, or cross-referencing multiple sources during research. Official documentation cites contract review as an example: let the agent handle one or two contracts directly, but use dynamic workflows for parallel reading when there are more.

To use it, set the multiagent type to multiagent_20261001 in the agent configuration (workflows are enabled by default under this type), then instruct Claude to "run a workflow." Claude handles both planning and scheduling. In Claude Code, you can get started immediately by typing /claude-api managed-agents-onboard bug-hunter. Templates are also available at https://t.co/UkKzzKVvfH.

Be mindful of costs. Every agent in a workflow consumes tokens, so running hundreds or thousands simultaneously causes bills to rise quickly. Anthropic recommends starting with a small-scope task and gradually increasing complexity. The documentation also suggests setting a "session budget" when creating a session; the workflow automatically pauses when the limit is reached and resumes after the quota is increased. Note that this budget can only be set at session creation and cannot be added later.

Agents temporarily defined within a workflow default to using the same model as the main agent. To delegate some tasks to cheaper models, you must create those agents separately beforehand and list them in the configuration.

↳ Quote @ClaudeDevs (ClaudeDevs): Claude Managed Agents dynamic workflows are now available in public beta. It's a new type of multiagent orchestration, built for your most ambitious workloads.

A lead agent writes a plan that runs across many agents in phases, combining the results at the end.