Title: SkillSmith: Co-Evolving Skills and Tools for Self-Improving Agent Systems

Abstract:

While recent advancements in self-evolving agents demonstrate that skills can be discovered, refined, and accumulated through execution, current frameworks face significant limitations. Most existing approaches assume a static tool layer and evaluate skills in isolation, which hinders their capacity to address tool-level failures or analyze inter-skill interactions. To overcome these challenges, we introduce SkillSmith, a framework designed for synergy-aware co-evolution of skills and tools.

SkillSmith establishes a unified proposal space where reflection generates atomic bundles capable of jointly modifying both skills and tools. This approach enables the system to wrap, edit, compose, split, or retire tools whenever the evolution of skills reveals a gap in reusable capabilities. To navigate this joint search space effectively, SkillSmith employs an ecological utility model based on Lotka-Volterra dynamics. By estimating an interaction matrix from execution traces, the framework captures the pairwise complementarity and conflicts among skills, thereby providing essential pressure signals for retrieval, prioritizing mutations, and determining which components to retire.

Additionally, SkillSmith maintains a record of anti-patterns—such as failure signatures, causal attributions, and remedies—to expedite diagnosis and reject proposals that repeat established errors. Our experiments across five Qwen3.5 model scales and three benchmarks, including WildClawBench, demonstrate that SkillSmith consistently surpasses strong baseline methods. Notably, performance gains become more pronounced as task complexity and the frequency of multi-skill co-activation increase.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC