Cleaning an AI agent's history can hurt when the world still holds the error
Tool-using language-model agents build up histories of their own mistakes. A system that compacts, repairs or hands off such a history decides what it says about those mistakes, while the effects of the mistakes stay in the world. The paper notes that prior work cleans the history at a fixed world, where showing a corrected step and saying it is done move together. Its question is whether the same correct content helps or hurts depending on how the history states it and what the world holds.
The method: the authors freeze an agent mid-task with its exact world state, edit either the history or the world, and let a receiver model continue from there.
On AgentCE-Bench, the agent first commits a wrong value. The world is held fixed, and the correct value is added by one of two natural edits: a truthful note beside the preserved error, or the agent's own completed action. The metric is recovery, the share of remaining targets the receiver gets right. The note raises recovery from 51% under a matched neutral note to 66%. The completed action lowers it below the unedited history, from 50% to 44%. So the correct value being present does not by itself predict the effect.
On AppWorld, the test follows a database write. The authors vary two things: whether the write's receipt reports success or failure, and whether its effect is actually present. A success receipt lowers whole-task success from 45% to 28% while the effect is absent. It raises whole-task success from 38% to 49% once the effect is present. The two effects differ by 27 points pooled over two receivers, a difference positive in six of eight receivers at 97.5% and seven at 95%.
The authors' conclusion is that a history edit's effect depends on the world it describes. A history made to look clean while the world still holds the error can leave an agent worse off than no edit at all.
Key facts
- On AgentCE-Bench, a truthful note beside the preserved error raises recovery from 51% (matched neutral note) to 66%.
- On the same benchmark, adding the agent's own completed action lowers recovery below the unedited history, from 50% to 44%.
- On AppWorld, a success receipt for a database write lowers whole-task success from 45% to 28% when the effect is absent, and raises it from 38% to 49% when the effect is present.
- The two receipt effects differ by 27 points pooled over two receivers, positive in six of eight receivers at 97.5% and seven at 95%.
- The authors conclude that a history edit's effect depends on the world it describes.
Why it matters
Agents that run long tasks collect their own errors in their history, and systems often compact, repair or hand off that history. Prior work cleaned the history at a fixed world, where showing a corrected step and saying it is done move together. This paper pulls the two apart and finds that the same correct content can help or hurt. A tidy-looking history is not safe by itself.
Who it affects
The paper frames the issue for any system that compacts, repairs or hands off a tool-using agent's history, because such a system decides what the history says about past mistakes while the effects of those mistakes stay in the world. The receiver model that continues the task is the one that pays when the history and the world disagree.
How to use it
This is a research finding, not a tool. The source states no practical recommendation beyond its final finding: a history edit's effect depends on the world it describes. Read that as a prompt to check how an edit states the correction, and what the world actually holds, before trusting the edit.
How solid is it
The numbers come from the paper's abstract and cover two benchmarks, AgentCE-Bench and AppWorld. The AppWorld gap is reported pooled over two receivers, positive in six of eight receivers at 97.5% and seven at 95%. The frozen-agent setup, with the exact world state kept, lets the authors change the history and the world separately. The source does not name the receiver models. It also does not say which AppWorld numbers belong to which receiver, or whether the 50% to 44% figure is pooled across receivers.
Risks and caveats
The authors say a history made to look clean while the world still holds the error can leave an agent worse off than no edit at all (note the wording: can, not always). The results rest on two benchmarks and a constructed freeze-and-edit setup. The source gives no sample sizes, task counts or confidence intervals beyond the stated receiver counts and confidence levels.
“a history made to look clean while the world still holds the error can leave an agent worse off than no edit at all”
— From the paper's abstract (arXiv 2611.14654)