📝Agents Managing Agents is a multi-agent arrangement in which supervisory agents spawn, assign, monitor, and correct other agents, so a person directs a hierarchy rather than each individual worker.
The pattern extends orchestration by one or more levels. A top agent breaks a goal into workstreams and launches agents to run them, and those agents may launch workers of their own. Supervisors track progress, answer workers' questions within their authority, retry failures, and escalate decisions they cannot make.
Hierarchy lets work scale beyond what one person can watch. A human reviews a handful of supervisor reports instead of dozens of sessions. Products expose this in forms such as 📝Claude Coordinator, the role 📝Claude plays in a Claude Code project, routing incoming tasks to parallel worker threads.
Each added layer multiplies cost and the risk of compounding errors. Effective hierarchies keep clear lanes so agents do not interfere with one another's work, reserve owner-only actions for the human, and require handoff notes so state survives when a session ends.
My development stack runs three layers deep: an Orchestrator session I start from the app or iOS directs Coordinator shifts, and each Coordinator launches Contributor runs. Owner-only actions come back to me as asks, so I steer the top of the tree instead of every branch.