AI

Cleaning an Agent's History Hurts When the World Still Holds the Error

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 13/11/2026, 07:48:24 UTC.

Briefings

Research on tool-using AI agents shows that cleaning their error histories can worsen performance if the world still contains errors. On AgentCE-Bench, adding a correct note to an agent's history increased recovery from 51% to 66%, but adding the agent's own completed action reduced it to 44%. On AppWorld, a success receipt lowered overall success from 45% to 28% while its absence raised it from 38% to 49%. These findings indicate that a history's effect depends on the current world state, and cleaning a history while the world holds an error can leave an agent worse off than no edit.

Key points

• Adding a correct note to an agent's history increased recovery from 51% to 66% on AgentCE-Bench.
• Adding the agent's own completed action to its history reduced recovery from 50% to 44% on AgentCE-Bench.

Context

Tool-using language-model agents accumulate histories of their own mistakes. Prior work cleans history at a fixed world, but this study asks whether the same correct content helps or hurts depending on how the history states it and what the world holds.

Limitations

Data is from arXiv preprint 2611.14654v1, which is unverified. Results are limited to specific benchmarks (AgentCE-Bench, AppWorld) and do not generalize to all environments or agent types.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 13/11/2026, 07:57:04 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiad36m3xzvufovv55wdjsxxnynhzp7hig27szql6dhn5dsjnwhjom