Agent handoffs are where multi-agent systems stop being routing diagrams and start becoming production software.
Handoffs between agents in a multi-agent system, i.e., how a customer is being transferred from one agent to another, typically a specialist to another specialist for another question, in practice are not trivial and affect the trace of the conversation of that customer. In particular, the first handoff does go wrong: billing believes that support is still answering the same question, support believes that billing has taken over the conversation with the customer, and fraud learned the whole story of the customer and now needs the approval status from the first agent.
The hard part of a multi-agent system is the handoff. Handoff contracts make ai agent orchestration into runtime software for buying.
Ownership moves, or it does not
Orchestration, as described by OpenAI, makes an important distinction between handing over a customer to a specialist continuing a conversation and using agents as tools a manager uses to arrive at a final answer while that manager retains responsibility for the answer to the customer’s question (OpenAI describes that split directly). This is what needs to concern buyers of a multi-agent system: is responsibility for a customer’s query being transferred to a specialist, or is a manager retaining responsibility for a customer’s question even while using agents to gather information and do work?
