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

SkillSmith introduces a co-evolution framework where AI agent skills and tools develop together rather than independently, using ecological dynamics to model skill interactions and anti-pattern tracking to prevent repeated failures. The system demonstrates consistent improvements across multiple benchmarks and model scales, particularly as task complexity increases.

SkillSmith addresses a fundamental limitation in current self-improving AI systems: the assumption that tools remain static while skills evolve. Traditional frameworks treat skill development and tool design as separate processes, preventing agents from identifying and fixing tool-level problems or understanding how skills complement or conflict with each other. This research demonstrates that joint optimization produces measurable performance gains, suggesting that future autonomous systems require more integrated architectural approaches.

The framework's ecological utility model, inspired by Lotka-Volterra population dynamics, provides an elegant solution to a complex coordination problem. By estimating interaction matrices from execution traces, SkillSmith identifies which skills work synergistically and which create conflicts. This mirrors biological coevolution—where predators and prey shape each other's evolution—applied to artificial systems. The anti-pattern recording mechanism adds practical value by allowing agents to learn from failure modes rather than rediscovering them repeatedly, accelerating the debugging cycle.

For the AI development community, SkillSmith suggests that scalable agent systems require ecological thinking: understanding how capabilities interact and compete for computational resources. The consistent improvements across different model scales (including Qwen 3.5) indicate the approach generalizes beyond specific architectures. The performance amplification on complex, multi-skill tasks hints that as agents tackle increasingly sophisticated problems, co-evolution becomes not optional but essential for maintaining efficiency and reliability.

The research opens questions about how such frameworks scale to hundreds or thousands of skills, and whether the ecological model captures all relevant interaction types in complex domains.

• →SkillSmith enables tools and skills to co-evolve jointly, allowing systems to repair tool-level failures that fixed-tool frameworks cannot address.
• →An interaction matrix based on Lotka-Volterra dynamics identifies skill complementarity and conflict, guiding retrieval and mutation priorities.
• →Anti-pattern tracking prevents agents from repeating known failure modes, accelerating diagnosis and improving proposal vetting.
• →Performance gains amplify as task complexity increases and multiple skills activate simultaneously, suggesting scalability benefits.
• →The framework outperforms baselines consistently across three benchmarks and five model scales, indicating generalization beyond specific architectures.