Multi-Agent Orchestration

Every handoff is a context loss. Justify it.

The default assumption is that more agents means more capability. The actual outcome is more tokens, more latency, more variance, and harder debugging — unless the handoff buys something a single agent with tools cannot.

When to use this skill

• The user is proposing a multi-agent system (Crew, Swarm, Supervisor, A2A).
• The user has a multi-agent system and reports loops, drift, runaway cost, or "no one ever decides".
• The user is comparing "subagents" vs "tools" for the same workload.
• The user is using a framework's multi-agent abstraction without a reason that survives "why not a tool?".

The handoff test

Before splitting work into a second agent, the handoff must clear all three: