SkillFoundry

Building Self-Evolving Agent Skill Libraries from Heterogeneous Scientific Resources

Shuaike Shen*, Wenduo Cheng*, Mingqian Ma, Alistair Turcan, Martin Jinye Zhang, Jian Ma†

Ray & Stephanie Lane Computational Biology Department, School of Computer Science, Carnegie Mellon University

[*Equal contribution · †Correspondence: jianma@cs.cmu.edu]

     

---

Overview

Modern scientific ecosystems are rich in procedural knowledge — repositories, APIs, scripts, notebooks, documentation, databases, and papers — yet much of this knowledge remains fragmented and difficult for agents to operationalize. SkillFoundry bridges this gap with a self-evolving framework that converts heterogeneous scientific resources into validated, reusable agent skills.

_{Figure 1. SkillFoundry framework overview: from domain knowledge tree to validated skill library.}

Key Results

267+ skills	mined across 28 scientific domains and 254 subdomains
71.1% novelty	vs. existing skill libraries (SkillHub, SkillSMP)
5/6 datasets improved	on MoSciBench benchmark
Genomics boost	substantial gains on two challenging genomics tasks

---

How It Works

SkillFoundry uses a domain knowledge tree as both a search prior and the evolving structure being updated, turning open-ended skill collection into a closed-loop acquisition process:

Step	Stage	Description
1	Tree Construction	Build a rooted tree where internal nodes are domains/subdomains and leaves are actionable skill targets
2	Resource Mining	Select focus branches and retrieve relevant resources (repos, APIs, papers, notebooks, databases)
3	Skill Compilation	Extract operational contracts and compile into reusable skill packages with metadata, dependencies, and tests
4	Multi-Level Validation	Apply execution testing, system testing, and synthetic-data testing
5	Tree Expansion	Insert validated skills as new leaves, expanding domain coverage
6	Refinement & Loop	Revise, merge, or prune failing/redundant skills; repeat from step 2

---

Repository Structure

SkillFoundry/
├── skillfoundry/             # Core automation framework (Python package)
│   ├── cli.py                #   CLI entry point
│   ├── orchestrator.py       #   Skill automation orchestrator
│   ├── campaign.py           #   Long-running campaign runner
│   ├── evaluation.py         #   Hierarchical skill evaluation
│   └── ...
├── scripts/                  # Utility & validation scripts
├── registry/                 # Taxonomy, resource registry, skill index
├── skills/                   # Reusable skill folders grouped by domain (27 domains)
├── tests/                    # Test suites (smoke, integration, regression)
├── site/                     # Generated project page (static HTML/JS/CSS)
├── ref/                      # Reference materials
└── Makefile                  # Build, validate, test, and smoke targets

---

Getting Started

Prerequisites

• Python 3.10+

Installation

git clone https://github.com/ma-compbio-lab/SkillFoundry.git
cd SkillFoundry
pip install -e .       # Install the skillfoundry package

Quick Validation

make validate        # Validate repository structure
make build-site      # Build the project page
make test            # Run unit tests

---

Framework Usage

The skillfoundry package provides a CLI for automated skill discovery, compilation, and evaluation. It orchestrates the closed-loop tree_check -> resource_search -> skill_build -> skill_test -> refresh pipeline.

Status

Inspect the current repository summary and identify high-value frontier leaves:

python3 scripts/sciskill_framework.py --json status --focus-limit 10

Cycle

Run one or more automation loops to discover and build new skills:

# Single loop
python3 scripts/sciskill_framework.py cycle --loops 1 --verification-mode standard

# Parallel workers with custom focus
python3 scripts/sciskill_framework.py cycle \
  --loops 2 --focus-limit 12 --stage-workers 4 \
  --stages tree_check,resource_search,skill_build,skill_test,refresh \
  --extra-context "Prioritize uncovered leaves in robotics and physics."

Design Skill

Design a skill from a specific task description:

python3 scripts/sciskill_framework.py design-skill \
  --prompt "Design a skill for literature-backed pathway enrichment benchmarking." \
  --verification-mode validate

Evaluate Skills

Run hierarchical evaluation (correctness repair, benchmarking, novelty checking):

# Single skill
python3 scripts/sciskill_framework.py evaluate-skills \
  --skill-slug openalex-literature-search \
  --verification-mode validate

# Full library
python3 scripts/sciskill_framework.py evaluate-skills --all --verification-mode none

Campaign

Run a long checkpointable campaign targeting specific domains:

python3 scripts/sciskill_framework.py campaign \
  --focus-term genomics --focus-term proteomics \
  --max-iterations 100 --max-runtime-minutes 450 \
  --stage-workers 6 --evaluation-workers 6

---

Citation

Citation information will be available once the paper is published. Check back later.

---

License

This project is licensed under the Apache License 2.0 and developed at Ma Lab, Carnegie Mellon University.