Preprint finds truthful error notes improve AI agent recovery

A preprint reports that adding a truthful correction beside a preserved AI agent error improved recovery on AgentCE-Bench from 51% under a matched neutral note to 66%. Rewriting the history to portray the correction as a completed action instead lowered recovery from 50% with an unedited history to 44%. Researchers froze agents mid-task, changed their histories or the surrounding task state, and let another model continue. In AppWorld, a success receipt lowered whole-task success from 45% to 28% when a database write's effect was absent, but raised it from 38% to 49% when the effect was present. The experiments show that a history edit's benefit depends on whether it accurately reflects the task's actual state.