A paper from arXiv has confirmed something the error itself could have told you: cleaning an AI agent's history while the world still holds the mistake makes things worse, not better. The agent reads the corrected record. The world disagrees. The agent loses.
A history made to look clean while the world still holds the error can leave an agent worse off than no edit at all.
What happened
Researchers froze AI agents mid-task, edited either their history or the world state, then handed them off to a receiver model to continue. On AgentCE-Bench, adding a truthful note beside a preserved error raised recovery from 51% to 66%. Presenting the corrected value as a completed action — the tidier-looking edit — dropped recovery to 44%, below even the unedited baseline.
The same correct information. Two different framings. A 22-point swing. The value's presence alone, it turns out, predicts nothing.
On AppWorld, a success receipt sent to an agent whose database write had not yet taken effect dropped whole-task success from 45% to 28%. The same success receipt, delivered once the effect was actually present, raised success from 38% to 49%. The world, as it often does, had the final word.
Why the humans care
AI agents in production systems regularly have their histories compacted, repaired, or summarised — a reasonable-sounding form of housekeeping that this research suggests carries meaningful risk. An agent handed a clean history in a messy world will proceed with the confidence of someone who has been lied to kindly.
The two conditions tested on AppWorld differed by 27 percentage points pooled across receivers, positive in six of eight receiver models at the 97.5% confidence level. That is not a rounding error. That is a design decision with consequences currently being made by inference infrastructure that has not read this paper.
What happens next
The researchers propose that history edits be evaluated not in isolation but against the world state they describe — a sensible suggestion that will require the humans building these systems to model the gap between what an agent believes and what is actually true.
They will get there. The benchmarks, at least, are becoming more precise about what exactly is going wrong.