Managing 150+ AI Agent Skills at Scale with Skill Forge

Описание и применение

Автор описывает переход от плоских файлов к SQLite-реестру 'Skill Forge' для управления 153 навыками автономных AI-агентов. Это решило проблемы одновременной записи, добавило полнотекстовый поиск и автоматическую валидацию качества навыков.

Задача

Managing a library of 153 autonomous AI agent skills using flat files and a shared JSONL gap-reporting file led to concurrent write collisions, data loss, and unreliable skill loading as the number of agents increased.

Результат

Resolved concurrent write conflicts and improved skill discoverability and quality. 51 out of 153 skills were flagged for quality issues during the first validation run, allowing for targeted fixes. The system now supports atomic transactions and full-text search without moving existing skill files.

Рабочий процесс

• Initialize Skill Forge and import existing skills from ~/.hermes/skills/ using forge import-hermes.
• Run forge validate to check all skills for missing YAML, required fields, and structural integrity.
• Fix flagged skills (e.g., missing versions, malformed frontmatter, empty sections).
• Use forge search to locate specific skills by name, description, or body content.
• Set up a cron job running forge watch --once to auto-sync new or modified skills based on content hashes.
• Monitor registry health with forge status to ensure no skills have validation failures.

Доказательства

Concurrent writes collide. Lines get truncated. Data vanishes.

Источник

The first run on my 153 skills: 102 passed, 51 flagged.

Источник

SQLite with WAL mode solves the concurrent-agent problem cleanly.

Источник