AI agent for code documentation generation

Why they built it

Documentation rots quickly due to code changes, making onboarding painful and LLMs ineffective on large codebases without specialized analysis

What worked

YAML templated prompts enabled collaboration; structured agent outputs created persistent AI memory; GitLab cron automated updates

What broke or was painful

CrewAI caused prompt tampering, parsing errors, incomplete tasks, poor observability; switched to pydantic-ai

The result

50+ repos auto-documented weekly; onboarding reduced from weeks to 1 day; 639 GitHub stars