Using CI pipelines as agent loops for AI-assisted coding is explored through a practical setup where an AI agent iterates on a pull request until all pipeline checks pass. The approach involves building a longer, stricter-than-usual CI pipeline with orthogonal, deterministic steps: compilation, secrets scanning (Gitleaks/Trivy), linting (ESLint/Biome), unit/integration/E2E tests, schema migration checks, dependency security scanning (OSV-Scanner), quality checks (SonarQube, cognitive complexity), container builds, infrastructure security, and a final LLM code review step. Three guiding principles are outlined: orthogonal steps, deterministic checks, and linear ordering with faster checks first. The agent drives fixes via the GitHub CLI, iterating until the pipeline is green, freeing developers to focus on higher-level concerns.
