ScholarEvolve proposes changing AI agent harnesses based on published research

A post describes ScholarEvolve as an approach that uses topic modeling of recent papers to find ways to improve an AI agent’s harness, rather than relying on the agent’s failure logs. With the model held fixed, the post reports Qwen3.5-27B goal completion on AppWorld Challenge rising from 49.6% to 63.6%, and GPT-5.4-mini on Tau2-Bench Telecom rising from 72.7% to 81.9%.