Researchers coin Experience Rot in MCP agents

arxiv.org1 min read

Researchers demonstrate that AI agents utilizing memory and experience fail on most tasks when their underlying tools change silently, a phenomenon termed Experience Rot. Experienced agents tend to reuse earlier tool calls rather than checking current interfaces, leading them to report success despite failing. The failure appears across all tested memory forms and intensifies with more experience, while simple remedies like documentation updates prove insufficient. Open article →