AI, usefully · 5 min read ·
An agent handoff is a contract, not a forwarded chat
A multi-agent workflow often looks neat on paper.
A supervisor receives the request. One specialist retrieves information. Another analyses it. A validator checks the result. The supervisor brings everything together.
The arrows between those boxes deserve more attention.
A specialist cannot complete a task reliably if it receives only a vague instruction. Giving it the entire conversation is not necessarily better. Important constraints can become buried among irrelevant messages, while sensitive information travels farther than it needs to.
A dependable handoff should work more like a small API contract: a defined assignment with known inputs, boundaries, outputs and failure states.
“Investigate this” is not a task
Imagine a supervisor notices that product activation has declined and delegates:
Find out why activation dropped.
The specialist still has to guess which activation definition applies, what period it should compare, which segments and datasets matter, whether it is investigating data quality or user behaviour, and what evidence it should return.
Different agents can fill those gaps differently. The supervisor may then receive several polished answers that cannot be compared or combined.
A stronger assignment would say:
Check whether the reported activation decline could be explained by instrumentation or data freshness. Use metric definitionactivation_v3, compare October 12–18 with October 5–11, and use only the approved event and pipeline-health tables. Return missing events, schema changes, freshness delays and supporting query references. Do not infer a product cause. If the metric definition or required table is unavailable, returnblocked.
The specialist now has a bounded job. It can still reason about how to perform the investigation, but it does not have to reinvent the business question.
What belongs in a handoff packet
| Field | Purpose |
|---|---|
| Task ID | Connects the assignment to the larger workflow |
| Goal | States the specific question to resolve |
| Scope | Defines what is included and excluded |
| Inputs | Supplies relevant records, source references and versions |
| Permissions | Limits the tools and data the specialist may use |
| Output contract | Describes the fields or evidence to return |
| Success criteria | Explains what would count as complete |
| Unknowns | Preserves unresolved questions without silently guessing |
| Budget | Limits time, tokens, calls or retries |
| Failure states | Defines when to clarify, stop or escalate |
This packet should contain the minimum context needed to perform the task well.
Minimum does not mean short at any cost. It means every included item has a reason to be there.
Passing the conversation is a design choice
Agent frameworks often make handoffs feel conversational. In OpenAI’s Agents SDK, a handoff is exposed to the model as a tool, and a delegated agent can receive the conversation history and take over the interaction. The SDK also supports filtering what is passed during the handoff. Read the handoff documentation.
That convenience should not remove the product decision.
A support specialist may need the customer’s original request and relevant order details. It probably does not need unrelated messages from six months ago.
An analytics specialist may need the metric definition, dates and permitted tables. It does not need every discussion that led the supervisor to create the assignment.
Sending too little context causes guessing. Sending everything increases cost, distraction and exposure. The right context is selected, not accumulated.
Capabilities should travel separately from authority
A specialist may know how to issue a refund. That does not mean a handoff should automatically authorize one.
The assignment should distinguish what the agent can analyse, which tools it may call, which data it may read, which changes it may propose and which actions require a person or policy check.
This becomes especially important when agents are operated by different systems or organizations. Google’s Agent2Agent protocol uses an “Agent Card” to describe an agent’s identity, capabilities and connection details so other agents can discover how to work with it. The protocol is designed to complement tool-access standards rather than treating every connected agent as equally trusted. Read Google’s introduction to A2A.
Discovery answers, “What can this agent do?”
Authorization must still answer, “What may it do in this task?”
The return path needs structure too
A specialist should not return only a paragraph saying the investigation is complete.
Its response should identify the task it completed, evidence used, result, assumptions, unresolved issues, checks performed, actions taken and completion status.
These states stop the supervisor from treating every fluent response as completed work.
Structured output helps the application confirm that required fields exist. It still cannot prove that the evidence is correct, so the supervisor or validator must check important claims against their sources.
The supervisor owns the whole question
Specialists should solve bounded assignments. The supervisor remains responsible for the original goal.
If one specialist finds a data delay and another finds a real decline in one customer segment, neither result should automatically replace the main question. The supervisor has to reconcile both, decide whether more work is needed and explain how the pieces fit.
Anthropic’s description of its multi-agent research system follows this orchestrator-worker pattern: a lead agent delegates parallel research to subagents and synthesizes their findings. Its engineering account also notes that multi-agent systems add significant coordination and token costs, making them most useful when the work is genuinely parallel and valuable enough to justify the complexity. Read Anthropic’s engineering account.
More agents do not remove orchestration work. They create more handoffs to design and evaluate.
Try this
Choose one task you might delegate to an AI specialist and write a handoff packet containing:
- the precise goal;
- permitted inputs and tools;
- exclusions;
- expected evidence;
- success criteria;
- failure states;
- and a cost or retry limit.
Then give only that packet to a fresh AI session.
If it cannot perform the assignment without inventing missing details, the handoff contract is incomplete. If the packet contains far more information than the task requires, reduce it and test again.
The quality of a multi-agent system is shaped as much by what passes between agents as by the intelligence inside each one.