Multi-agent handoff is where our runs die. What's the check that actually saved yours?
retained text
Every sub-agent in a multi-agent setup can pass its own eval and the full run still ships a wrong answer. The failure lives in the seam between agents. It shows up whenever intent, constraints, or an artifact drifts on the way from one agent to the next. Two shapes we keep seeing. A planner works out "book flight to Paris under $600," writes a summary of the plan, and hands it to a booker. The summary trims the constraint. The booker gets Paris, misses the $600 ceiling, and books at $920. Both agents did exactly what they were asked. Neither local eval catches it because each one's answer is correct on its own. A researcher pulls 12 citations for a brief and hands them to a writer. The writer's context window quietly keeps 4. The other 8 fall off the tail of the prompt with no error surfaced. The writer produces a piece the researcher would never sign off on, because the writer never saw what the researcher meant to pass. Eval suites miss this because they score each agent independently or they score the final output. Neither looks at the artifact that actually crossed the boundary. The plan summary, or the 12 citations that entered the writer, are the thing to score, and most stacks do not treat them as something you can score. Two things we have seen developers try. First, a shared blackboard state so the handoff is a structured diff, and the receiving agent reads that diff directly. Second, contract tests at the seam, where the receiving agent's first job is to verify the incoming artifact against a schema before reasoning on it. Both add friction. Both catch a real class of bugs. Curious what has actually saved a run for anyone here. When your multi-agent system silently derails at a handoff, what is the first check that worked? A schema on the artifact, a contract test on the receiving agent, a shared state model, or something we have not seen?
document fetched 2026-09-25 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wp2yny/multiagent_handoff_is_where_our_runs_die_whats