Agent Handoff
The transfer of a task and its context from one specialised agent to another, and the point at which multi-agent systems most often lose information.
Multi-agent designs decompose a problem across specialised agents. The handoff is where the design succeeds or fails, because each transfer is a lossy summarisation.
What is lost: nuance from the original request, constraints established earlier, and the reasoning behind earlier decisions. By the third handoff the last agent may be working on a materially different problem from the one the user posed.
Design choices that determine how much is lost: what is passed — a summary, the full history, or a structured task specification (the last is usually best, because it is explicit about goal, constraints and what has been established); who decides the next step — a supervisor, or the agents themselves; and whether the handoff is one-way or returns.
The honest assessment: most multi-agent systems would work better as one agent with more tools, or as a deterministic workflow calling models at fixed points. The multiplication of cost, latency and failure modes is real, and the benefit is genuine only when sub-tasks need genuinely different tools, permissions or context that would not fit together.
Where it does apply — a research agent, a writing agent and a fact-checking agent, or per-domain specialists with distinct permissions — the operational requirements are end-to-end tracing across agents, a global iteration and cost budget rather than per-agent limits, and loop detection, since agents handing work back and forth is a common and expensive failure.