SkillAlchemy

Turn people, methods, and experience into installable, reusable agent skills.

English · 中文 · 日本語

SkillAlchemy is an open-world agent skill creation system that turns underspecified skill briefs and open-world sources into installable, reusable agent skills.

It discovers omitted requirements, identifies executable procedures from heterogeneous public sources, determines how broadly each procedure is justified by evidence, and compiles the admitted knowledge into agent skill packages.

🔥 News

• 2026-08-24 — Our paper, SkillAlchemy: Open-World Agent Skill Creation, is now available on arXiv.

Overview

Creating reliable agent skills is difficult when the target capability is unfamiliar: task descriptions omit important requirements, expert-written procedures may not exist, and execution traces may be unavailable.

SkillAlchemy approaches skill creation as a source-grounded procedure-admission problem.

Given an underspecified skill brief and access to open-world sources, it:

1. Discovers implicit requirements that are missing from the original brief.
2. Acquires grounded procedures from documentation, repositories, papers, issue reports, and other public sources.
3. Determines procedure scope — deciding whether evidence supports a reusable instruction, a scoped example, or exclusion.
4. Compiles an installable Skill package that agents can load and use directly.

Results

We evaluate SkillAlchemy on 87 tasks from SkillsBench v1.1 across four agent–model configurations.

SkillAlchemy achieves the highest overall task pass rate in 3 of 4 configurations.

Averaged across all configurations, SkillAlchemy reaches a 55.8% task pass rate:

• +19.9 percentage points over no-skill execution.
• +8.6 percentage points over the strongest automated skill-creation baseline.
• Performance comparable to human-curated skills, slightly exceeding them on average in our evaluation.