By Vilius Vystartas | May 2026 I run a lot of AI agents. Not chatbots — autonomous agents. Cron jobs that monitor my infrastructure every hour. Self-improvers that analyze past sessions and encode learnings. Delegated coders that build features while I sleep. Together they load from a library of 153 reusable skills — structured procedures that tell an agent how to do something specific, from sending iMessages to debugging SPFx builds. The system worked fine when I had 20 skills and one agent. It started breaking when the numbers climbed. The Problem That Forced My Hand Here's the setup: each skill lives as a SKILL.md file in ~/.hermes/skills/ . When an agent loads a skill and discovers it's broken, missing steps, or out of date, it records the problem in a shared skill_gaps.jsonl file. Later, I review the gaps and fix the skills. This is fine when one agent writes to the file at a time.…