Suppose an agent hands you a note saying a new account record was saved. You have two things you can check: what the note says and whether the record exists. An agent taking over the work faces that same choice.

Researchers at Salesforce AI Research paused agent runs and let a receiving model continue from a saved point. In one AppWorld experiment, they independently changed two things: whether a database operation’s receipt reported success and whether the database contained its effect. When the effect was absent, a success receipt lowered completion of the whole task from about 45% to 28%, pooled across two receiving models. When the effect was present, the success receipt helped: completion rose from about 38% to 49%. AppWorld experiment

A separate test kept the environment fixed and changed how the history described a wrong value. A note acknowledging the error and supplying the right value improved recovery over a neutral note. Rewriting the history as if the action had already succeeded performed worse than leaving it alone. These were controlled edits on two benchmarks. The authors didn’t test a complete handoff or history summarization policy in production.

What should a handoff call done? Give the next worker separate entries for what was tried and what the system confirmed. Leave an unresolved status visible when there’s no independent confirmation. After a database write, read the target record before continuing or retrying. The answer belongs in the handoff, where the next worker can find it.