Skills for AI agents: comparing four incompatible approaches

说明与用途

本文阐述了 AI 代理技能的四种不兼容方法，并提供了比较矩阵和静态审计脚本，用于评估来源、版本控制、权限和可执行内容。

问题

Choosing a skill framework for an AI agent is ambiguous because the term ‘skill’ can refer to four distinct architectures—instruction, trained, local package, or registry—each with different trust boundaries, update paths, and risks.

结果

Teams can systematically compare skill approaches, identify risks, and choose the most appropriate architecture. The audit script demonstrates how to catch flaws in a demo skill, ensuring safer deployment.

工作流

• 1. Identify the skill type (instruction, trained, local package, registry).
• 2. Gather provenance information (Git commit, training run, registry publisher).
• 3. Verify versioning (file diff, artifact hash, semver, pinned digest).
• 4. Assess permissions (agent tools, manifest declarations, publisher policy).
• 5. Check for executable content (scripts, binaries, remote execution).
• 6. Run the static audit script on the skill artifacts.
• 7. Review audit results and decide on the appropriate skill approach.

证据

"Add a skill to the agent" can mean four different things: write a Markdown instruction, retrain the model, drop a folder of scripts into the repository, or install something by name from a remote catalog.

来源

The four audit criteria are deliberately boring, because boring criteria are the ones that catch incidents: Provenance, Versioning, Permissions, Executable content.

来源

We will create a small local-package skill called changelog-summarizer. It has an instruction file, a manifest and one shell script.

来源