SKILLFOUNDRY: Building Self-Evolving Agent Skill Libraries from Heterogeneous Scientific Resources
Summary: arXiv:2604.03964v1 Announce Type: new
Abstract: Modern scientific ecosystems are rich in procedural knowledge across repositories, APIs, scripts, notebooks, documentation, databases, and papers, yet much of this knowledge remains fragmented across heterogeneous artifacts that agents cannot readily operationalize. This gap between abundant scientific know-how and usable agent capabilities is a key bottleneck for building effective scientific agents.
Introducing SkillFoundry
We present SkillFoundry, a self-evolving framework designed to convert diverse scientific resources into validated agent skills. These skills are reusable packages that encapsulate:
- Task scope
- Inputs and outputs
- Execution steps
- Environment assumptions
- Provenance
- Tests
How SkillFoundry Works
SkillFoundry organizes a target domain as a domain knowledge tree. The process involves several key steps:
- Mines resources from high-value branches of the knowledge tree
- Extracts operational contracts from the mined resources
- Compiles these contracts into executable skill packages
- Iteratively expands, repairs, merges, or prunes the skill library through a closed-loop validation process
Innovative Results
The framework has produced a substantially novel and internally valid skill library, with 71.1% of the mined skills differing from existing skill libraries such as SkillHub and SkillSMP. This innovation demonstrates that:
- Mined skills significantly improve coding agent performance across five of the six MoSciBench datasets.
- SkillFoundry can design new task-specific skills on demand that cater to concrete scientific objectives.
- The resulting skills lead to substantial performance improvements on challenging genomics tasks, specifically cell type annotation and the scDRS workflow.
Conclusion
Together, these results illustrate that automatically mined skills enhance agent performance on established benchmarks and domain-specific tasks. SkillFoundry expands coverage beyond traditional hand-crafted skill libraries, offering a practical foundation for developing more capable scientific agents. As the framework continues to evolve, it promises to bridge the gap between rich scientific knowledge and actionable agent capabilities, paving the way for more efficient and effective scientific research.