AI-Driven CI/CD: Orchestrating Agentic DevOps Pipelines

CI/CD pipelines are the nerve center of modern development, yet many steps (code review, test analysis, report generation) remain manual or tedious. We demonstrate how existing CI systems can themselves become agentic runtimes, plugging in LLM-powered agents as pipeline steps.

In this model, your CI tool (e.g. GitHub Actions, GitLab CI, Buildkite) continues to handle scheduling, containerization, and audit logs - and we simply insert an “AI agent step” wherever complex reasoning is needed.

This one-shot execution model is still multi-turn internally: an agent iterates (via ReAct or chain-of-thought) over files, tests, and tool calls all within a single pipeline job.

We cover practical patterns using tools like the open-source cicaddy framework and Buildkite’s agentic workflow components. For example, we show how a label on a PR can trigger a pipeline that launches a Claude-code agent to perform a first-pass code review (commenting issues on the PR), or automatically fix a broken build and submit a new PR with the patch.

Throughout, the talk is demo-driven: we’ll live-demo or show recordings of AI agents in action inside CI jobs. We highlight features like pipeline triggers, custom containerized agent steps, and token budget management. We also cover the dual benefits: saving developer time (e.g. AI-assisted code review) and improving code quality (e.g. catching hard-to-find bugs).

Security considerations are addressed too: since pipeline jobs run in isolated containers, agents only see the project workspace, not sensitive hosts or creds. The result is a powerful yet safe way to bring AI into standard DevOps workflows.