GitHub issues were piling up on my open-source project, and it was becoming a real burden.
I needed automation. My first attempt was n8n + LLM.
When an issue came in, the LLM would decide:
- Spam or not?
- Bug / feature request / question?
- Need more info?
- Can AI handle this directly?
Triage and auto-commenting worked surprisingly well.
Then I hit a wall.
The real core of this agentic workflow wasn't classification—it was the AI actually understanding code and taking action.
Reading code, leaving informed comments, making fixes and opening PRs... For any of that, the agent needs to run where the code lives.
That's when I landed on GitHub Actions.
Free compute for AI agents, and the GitHub integration is seamless.
I designed the architecture with Claude Code and documented all decision policies and role definitions in markdown.
Prototype done in 4 hours.
Next week: apply improvements, build test cases, deploy to the real project.
Curious to see how much time this actually saves the team.
What I'm experimenting with next:
- Structuring domain knowledge with Claude Skills (policies, metrics, builds, tests)
- Better AI Q&A interactions
- Test cases that validate LLM and agent behavior
Top comments (0)