Claude Orchestrates Multi-Agent Workflows with Up to 1,000 Sub-Agents

Notice: This article was created with AI.

What’s It About?

Anthropic has introduced a function for multi-agent orchestration on its Claude platform. In it, a superordinate agent takes over the planning and distributes tasks in real time to up to 1,000 specialised sub-agents. At most 64 of them can be active at the same time, while the workflow as a whole can run for up to 24 hours. The function is part of the Claude Managed Agents and is meant to make the handling of complex tasks more efficient through automatic division of work.

Background & Context

The new dynamic workflows allow a lead agent to break tasks down autonomously, activate additional sub-agents and consolidate their results. According to Anthropic, the platform can decide for itself when a multi-agent approach makes sense, provided the function has been explicitly switched on. During execution, progress is saved continuously, so that interrupted tasks can be resumed later.

In a benchmark test a code base was prepared with hidden defects. A single agent identified between 14 and 27 of 70 hidden code errors, while the multi-agent workflow consistently uncovered 66 errors. The sub-agents take on specialised partial tasks whose results the main agent collects and checks for contradictions or systematic patterns. This separation of planning and execution is meant to increase efficiency with complex problems.

What Does This Mean?

The orchestration allows extensive tasks to be handled in parallel by specialised agents without manual instructions on how to distribute the work.
The iterative verification by the main agent is meant to narrow down errors and raise the quality of the overall results.
With up to 1,000 managed sub-agents and long running times, the function is aimed at use cases in which large volumes of data or complex processes are to be processed automatically.
The automatic saving of progress makes uninterrupted workflows over longer periods possible.