WikiSkill separates cumulative experience from execution, enabling small models to outperform large ones through skill transfer

A new research paper introduces WikiSkill, a framework designed to address a central challenge in autonomous agent development: insights and experience gained through interaction and experimentation remain scattered across past optimisation logs rather than being accumulated systematically. The researchers propose an approach based on the co-evolution of agent skills and a persistent, wiki-style knowledge base, where raw execution experience is extracted and periodically integrated into a structured knowledge log that is subsequently used to update reusable code skills.

The proposed architecture relies ona systematic separation across three tracks: real-time execution experience, documented cumulative knowledge, and executable procedural skills. This separation prevents lessons learned from being lost across successive optimisation cycles, enabling the agent to refine its errors and consolidate successful solutions within the knowledge base. Benchmark tests across multiple models and environments showed that WikiSkill outperformed current methods in agent skill development, achieving a clear gain over baseline models operating without skill libraries.

The study highlighted two notable phenomena in agent architecture. First, skill accumulation complements model scaling, as larger models benefit more broadly from refined skills; however,smaller models equipped with advanced skills managed to outperform significantly larger models lacking those skills. Second, skills transferred effectively across different model families and architectures, with skills developed by an external model occasionally achieving higher performance than skills developed self-referentially. Ablation studies also confirmed that continuous knowledge accumulation within the wiki is the critical element for successful skill evolution.

These findings open an important practical path for technical teams and enterprises across the region, particularly in the Gulf, Egypt, and the Levant. Equipping smaller models with accumulated skills offers a highly efficient alternative that substantially reduces inference and cloud computing costs compared to relying exclusively on expensive large models. For developers and infrastructure leads, building an organised, persistent repository of agent experience allows organisations to run local or open-source models with an efficiency comparable to larger global systems, while preserving data sovereignty and maintaining the flexibility to switch providers without losing acquired skills.

This shift requires rethinking how enterprise agent systems are managed, moving from prompt engineering or costly retraining toward building a cumulative knowledge architecture that separates the record of experience from the inference engine.Investing in documenting agent experience and reusing skills provides technical architectures with exceptional flexibility, allowing teams with an independent knowledge and skill base to upgrade underlying models or migrate workloads across cloud environments without having to retrain agents from scratch, thereby increasing operational efficiency and reducing vendor dependence.