BONSAI Optimizes Frozen Agent Skills with Evolvability-Guided Search

When working with "frozen" AI agents whose weights cannot be updated, any new capability must be supplied through natural language "skills" – essentially, text instructions. The standard method for optimizing these skills, which accepts any edit that improves a held-out score, is limited because a single score cannot differentiate between a skill that is narrowly overfit and one that has broad potential for further improvement.This research introduces BONSAI, a new skill optimization framework that addresses this limitation by steering its search based on "evolvability." Evolvability refers to a region of document space's capacity to consistently produce viable variations under further mutation, a concept distinct from current fitness. BONSAI constructs skills using a Monte Carlo search tree where each child document is a mutation of its parent.The tree is traversed using an upper-confidence selection rule that combines a skill's individual fitness with the fitness of its mutational neighborhood. This design allows the mean score recorded beneath a node to estimate that neighborhood's evolvability without additional cost. This strategy focuses computational budget on areas that show continuous improvement, while also ensuring that currently weaker branches remain in contention for exploration. BONSAI ultimately delivers the best-scoring document found, demonstrating significant accuracy lifts over skill-free agents and other baselines.