Agent Skills
Agent skills are reusable capability bundles for AI agents: instructions, examples, scripts, and references that let an agent perform a specialized task more reliably than relying on the base model alone.
Sources in this batch
- The Agent Skills site presents skills as a standardized way to give agents new capabilities and expertise.
- The arXiv paper “Automating SKILL.md Generation for Computer-Using Agents via Interaction Trajectory Mining” studies whether skill libraries can be mined from GUI interaction trajectories. Its abstract reports that mined clusters can be readable on the source benchmark, but that readability did not yet translate into reliable transfer.
- Karpathy’s llm-wikis pattern is adjacent: both treat durable, inspectable artifacts as a way to compound agent competence.
Research interest
The surprising angle is that skill libraries may be inspectable without being transferable: the arXiv source reports readable mined clusters but weak downstream gains. For a CS researcher, this is a useful negative result because it separates human-legible decomposition from policy-improving abstraction, raising questions about representation, segmentation boundaries, and offline reward design.
Open questions
- Which parts of a skill are most transferable: procedural steps, examples, validation scripts, or environment-specific caveats?
- Can skills be mined automatically without hardening spurious behavior from the source trajectories?
Related
Batch 21-100 update
New related page: coding-agents-and-skills. The second batch adds evidence around skills-versus-docs evals, coding-agent components, Claude Code style workflows, GEPA/DSPy optimization, and open terminal/tool environments.
Updated: 2026-06-27