Anthropic Launches Public Beta for Claude Dynamic Workflows Supporting Thousands of Agents

Anthropic has opened the public beta for "Dynamic Workflows" in Claude Managed Agents, where the main agent first writes a plan, then assigns tasks to a large number of agents for parallel execution, and finally aggregates the results, allowing up to 1000 agents to be scheduled in a single run.

Claude Managed Agents is a managed agent service by Anthropic for developers, where agents run directly in the Anthropic cloud environment after developers configure models, tools, and prompts, without needing to build their own loops and sandboxes, supporting long tasks from several minutes to several hours. Previously, it supported a sub-agent mode, where the main agent assigns tasks and reads reports one by one, with a maximum of 25 agents concurrently in a single session.

The dynamic workflow mechanism is different: the main agent writes a program that the server executes in the background in stages, with context and results passed between agents through the program, allowing the main agent to continue conversing with the user and report progress at any time. In a test with a codebase of 116,000 lines embedded with 70 bugs, a single agent found 14, 15, and 27 bugs in three runs, showing significant result variability; however, the workflow found 66 bugs in each of its three runs.

This feature is suitable for tasks that can be broken down, including auditing large codebases, bulk code migration, reading hundreds of contracts, and cross-referencing multiple sources during research. The official documentation provides examples where one or two contracts can be handled directly by agents, and workflows are initiated for parallel reading as the number increases. To enable this, set the multiagent type in agent configuration to multiagent_20261001 (workflows are enabled by default under this type), and instruct Claude to run the workflow, with planning and scheduling handled by it. Relevant commands can be entered directly in Claude Code, and templates are also provided on the platform.

Each agent consumes tokens, and costs rise rapidly when running hundreds to thousands in parallel. Anthropic recommends testing with small-scale tasks before gradually increasing complexity. The documentation suggests setting a "session budget" when creating a session, which automatically pauses the workflow upon reaching the limit, and can be continued by raising the budget; this budget can only be set at creation and cannot be added later. Temporarily defined agents within the workflow default to using the same model as the main agent, and if cheaper models are needed for some tasks, those agents must be created separately and included in the configuration in advance.

This feature is driven by product capability expansion and developer adoption behavior. The flow of funds mainly comes from token fees paid by the calling party, directed towards Anthropic's cloud computing resources. Beneficiaries include enterprise developers needing to handle large-scale parallel tasks (improving efficiency through stability and scale), and Anthropic (increasing revenue through high concurrent usage); the pressure is on cost-sensitive users and the unstable performance of single-agent solutions on complex tasks.

Source: Public Information