Preprint reports graph-based instructions improved long-running AI-agent reliability

A preprint reports that a graph-based system for updating an AI agent's persistent instructions raised strict reliability on a held-out telecom task from 0.091 for Gemini 2.5 Flash without adaptation to 0.673, plus or minus 0.136. The proposed GRACE method stores instructions as a typed semantic graph, checks changes near the affected nodes and converts accepted updates back into deployment text. Across five replications, the final result also exceeded a flat-text updating baseline of 0.191, plus or minus 0.051. The findings come from one controlled harness and have not undergone peer review.