What is Multi-agent orchestration?

There is no single formal specification for multi-agent orchestration, so the cited page is one vendor's description. In Anthropic's account of its research feature, a lead agent analyzes the query, plans an approach, and spawns subagents that explore different aspects at the same time. The subagents search with tools, judge what they find, and return results to the lead, which synthesizes them and decides whether more research is needed. A separate agent then adds citations to the report.

The post reports that this setup beat a single agent on Anthropic's internal research evaluation and works best for broad questions that split into independent directions. It also states costs and limits. Agents use about four times the tokens of chat and multi-agent systems about fifteen times, so the tasks need to be valuable enough to justify it. Work that needs shared context or has many dependencies between agents, such as most coding, was described as a poor fit at the time.

The authors list practical lessons: task descriptions given to subagents must be detailed or agents duplicate work, early versions spawned far too many subagents for simple queries, errors compound in long runs, and debugging needs production tracing. Their lead agent also ran synchronously, so it could not steer workers mid-task.