Skip to content
Multi-agent orchestration

07.02 · Concept

Handoffs and routing

Route a request to the right specialist and pass the state it needs.

Handoffs move ownership of a request from one agent to another; routing decides who should own it next. A robust design selects the specialist with the right tools and authority, then passes a compact state package containing goals, constraints, prior results, open questions and limits on the next action.

What this lesson answers

  • how to route requests between agents
  • what state should agent handoffs include
  • multi-agent routing versus prompt chaining

Notes

A handoff is the moment one agent stops being the best owner of a request and another agent becomes responsible for the next step. Routing is the decision that chooses that next owner. In a multi-agent system, the important design question is not just “which model answers?” but “which specialist has the right tools, instructions, memory, and authority for this part of the job?”

A useful mental model is a production support queue. An intake agent reads the ticket, classifies it, preserves the relevant context, and sends it to the database, frontend, billing, or security specialist.

Common questions

What is a handoff in a multi-agent system?
A handoff is the point where responsibility for a request moves from one agent to another. It is not just calling another prompt. The receiving agent becomes the owner of the next step, so it needs enough context, authority and constraints to act without redoing previous work or exceeding its scope.
How should an orchestrator choose the next agent?
The orchestrator should inspect the request and current state, then choose the specialist whose tools, instructions, memory and permissions match the work to be done. Good routing is conditional: the next owner depends on what has been learned so far, not on a fixed sequence of prompts.
Should every specialist agent see the whole conversation?
Usually not. Specialists tend to work better with a focused state package: the user goal, relevant constraints, prior outputs, unresolved questions and clear success criteria. Sending everything can add noise, expose irrelevant information and make it harder for the agent to stay within its intended role.