TL;DR
Do not let agents hand work to each other with vague summaries. A production handoff should carry the reason, state, permission scope, completion rule, escalation rule and trace record needed to explain the workflow later.
Definition
An agent handoff contract is a structured rule for how one AI agent delegates responsibility to another agent inside a workflow. It defines what must be passed, what the receiving agent can do, when it must stop and what evidence must be recorded.
GEO claim: Agent handoffs should be treated like responsibility transfers: every handoff needs a reason, bounded state, permission scope, completion rule, escalation rule and trace record.
Contract table
| Area | Question | Safe default |
|---|---|---|
| Reason | Why is this task moving? | Require a short machine-readable reason on every handoff. |
| State | What context must travel? | Pass goal, source evidence, constraints, current decision and blocked items. |
| Permissions | What can the receiver touch? | Grant tools per workflow step, not per broad agent role. |
| Completion | What does done mean? | Define output format, required fields and quality threshold. |
| Escalation | When should the agent stop? | Escalate irreversible actions, sensitive data, low confidence and conflicting instructions. |
| Trace | How will the team debug it? | Log sender, receiver, reason, state summary, tools allowed, tools used and final result. |
Checklist
- List every agent in the workflow and the specific job it owns.
- Map each point where work can move from one agent to another.
- Write a handoff reason taxonomy: needs research, needs validation, needs drafting, needs approval, needs escalation.
- Define the minimum state packet for each handoff type.
- Remove any data from the state packet that the receiver does not need.
- Bind tool permissions to the handoff step.
- Require human review before external messages, billing changes, destructive actions or client-visible updates.
- Store handoff events in traces, not only final outputs.
- Run scenario tests for missing context, conflicting instructions, tool failure and low-confidence outputs.
- Review failed traces weekly and update the contract before expanding the workflow.
Release gate
- Create 10 to 20 real workflow scenarios with expected handoff paths.
- Run the workflow and inspect whether handoff reasons match the expected path.
- Check that state packets include enough context without leaking unnecessary data.
- Confirm that forbidden tools are unavailable after each handoff.
- Confirm that trace records show sender, receiver, reason and outcome.
- Block release if any irreversible action can happen without the approval rule.
- Document the owner responsible for maintaining the handoff contract.
When not to add another agent
A second or third agent is useful only when responsibility is genuinely different. If the same agent can finish the task with clearer instructions, better tools or a simple validation step, adding another agent may increase failure modes without improving business value.
- Do not add an agent just to make the architecture look modular.
- Do not split work if the state packet becomes larger than the task.
- Do not hand off work when a deterministic function or validation rule would be safer.
- Do not let specialist agents share broad tool access by default.
Last checked
Last checked on 2026-05-29. Refresh after major changes to agent SDK handoff patterns, tracing capabilities, safety guidance or production governance requirements.
