Feedback Dynamics in Self-Evolving Agent Skills: A Controlled Study

A recent study published on arXiv (2608.02636) explores the factors that enhance AI agents through self-evolving skill systems. The research establishes a controlled evaluation framework utilizing five benchmarks and three models, executing 42 feedback runs across 14 model-benchmark combinations. In each scenario, parameters such as executor and optimizer configurations, revision methods, validation rules, and round budgets remain constant, while the feedback provided to the optimizer varies: encompassing successes and failures (Normal), failures alone, or successes alone. Findings indicate that evolution is infrequent, with only 55 out of 388 candidates achieving distinct validation bests. Validation-based selection favors an evolved skill in 11 out of 14 scenarios, highlighting the significant role of feedback dynamics in self-improvement. The paper tackles critical inquiries regarding the conditions that promote further evolution, the impact of successful and unsuccessful trajectories on revisions, and whether additional computation at test time can yield similar benefits. These insights are crucial for developing autonomous agents capable of learning from experience without altering their foundational models.