Skip to main content

Hierarchical multi-agent systems

AI concepts

A pattern where a supervisor (or orchestrator) decomposes work and delegates to specialist workers—each may have its own tools and prompts—then synthesizes results.

Hierarchical setups reduce overload on a single model: a coordinator handles routing and integration while workers focus on domains (logs, metrics, code search, etc.).

Design tensions

You must decide how workers communicate (full traces vs. summaries), who may spawn sub-tasks, and what happens when a worker fails (retry, substitute, escalate). These choices mirror real engineering incident workflows.

See also

Plan-and-Execute and ReAct.