Large language model agents can improve their capabilities by evolving reusable natural-language skills from execution experience. This process lets agents accumulate practical knowledge over time, turning raw interactions into structured, transferable skills.
In real-world deployments, however, this experience is distributed across data owners who cannot pool their data. That constraint makes it difficult to apply conventional centralized training or skill-sharing approaches, since the raw experience cannot be brought together in one place.
The new framework, FedSkillGit, addresses this by borrowing ideas from version control. It enables federated skill evolution, allowing agents to develop and share skills across distributed data owners while respecting their inability to pool data. The approach is described in a recent arXiv preprint.