Agent Handoff — AI Glossary

Reference details

Related terms

• Multi-Agent System
• Orchestrator Agent

How it works

1. The sending agent packages a context bundle: the sub-task, what has been established, intermediate results, and any constraints.
2. What to include is the design decision — too little and the receiver re-derives work, too much and it inherits irrelevant context that dilutes its focus.
3. The receiving agent starts from that bundle rather than from the full history, which is what keeps each agent's prompt small enough to behave predictably.
4. Control transfers explicitly, so exactly one agent owns the task at a time and there is no ambiguity about who is acting.
5. Results return through the same boundary, and an orchestrator checks them against the sub-task before accepting them into shared state.

Example: The cost of an under-specified bundle

A research agent establishes three constraints over eight turns, then hands off to a writing agent with only the topic. The writer produces output violating two constraints, which review catches, so the work is redone — the eight turns of establishing context bought nothing, and the system paid for the writing twice. Including the constraints costs perhaps 200 tokens in the bundle. The asymmetry between 200 tokens and a repeated generation is why bundle design matters more than it sounds.

Commonly confused with

Used interchangeably with delegation and with routing, which differ in what transfers. Routing picks a destination for a request and ends there. Delegation gives a sub-task to a worker and expects a result back, with the delegator retaining ownership. A handoff transfers control itself, so the receiver owns the task from that point — which is why the context bundle has to be complete. Conflating them produces the common bug: an agent handed control without the context to exercise it, which then restarts work already done.

When to use it

Reach for it when:

• Control must pass between specialised agents, each with its own narrow prompt.
• Keeping per-agent context small matters for predictable behaviour.
• The receiving agent needs established facts without the whole transcript.

Reach for something else when:

• Only a destination needs choosing, which is routing.
• The sender should retain ownership and receive a result, which is delegation.
• The bundle cannot be made complete, where the receiver will redo work.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Agent Handoff — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path