How one engineer and an AI agent built a CI system that tested, corrected, and extended itself — feature by feature
I started with an empty repository and a question: could one engineer and an AI agent build a production-grade CI system from scratch?
Not a toy. A real system — multiple microservices in Go, DAG-based parallel pipeline execution, webhook-driven builds, AI code review, security scanning, and a full MCP server that any AI agent could call programmatically. Something conceptually similar to Jenkins, but designed for a world where AI agents are first-class participants in the development process.
The system is called Relay CI. Building it would have previously required a dedicated team and months of runway. I built it with Claude Code in a fraction of that time.
But the more interesting story is not that it got built fast. It is what happened partway through — when the system the AI was building became the system that validated the AI’s own output. The tool and the toolmaker merged into a single loop.
130 Lines of Intent
The single most consequential decision I made was not about technology. It was writing a 130-line document called CLAUDE.md and checking it into the root of the repository.