Multi-Agent Routing: Who Does What?

A growing note

This is the first seed in the Agentic AI section — planted early so the garden has its shape. It will grow.

One capable AI agent is useful. A team of specialists — a researcher, a coder, a reviewer, a planner — is far more powerful. But the moment you have a team, you inherit the question every team faces: who does what? The machinery that answers it, task by task, is called routing, and it quietly determines whether a multi-agent system feels like a well-run workshop or a room full of people talking over each other.

Three routing patterns

1. Orchestrator and workers. A lead agent reads the incoming task, breaks it into pieces, and delegates each piece to a specialist. The orchestrator owns the plan and assembles the results. Most capable pattern; costs an extra model call per hop.

2. Classifier routing. A small, fast model tags each request — billing, bug report, research question — and forwards it to the matching specialist. Cheap and predictable; ideal when the categories are known in advance.

3. Handoffs. Agents pass the conversation directly to a peer when they recognize the work is outside their lane, carrying the context along. Flexible and decentralized; needs guardrails so tasks don’t ping-pong.

What separates good routing from chaos

• The specialization gain must beat the coordination cost. If one agent could do the whole job well, routing is overhead.
• Every agent needs a contract. Clear inputs, clear outputs, so the router can compose them without guessing.
• The system needs a feedback loop. Watch where tasks fail or bounce, and reroute. Stable systems correct themselves — stars figured that trick out billions of years ago.