Headless Mode, CI Automation & the Agent SDK

Every other subject in this track assumes a human is watching — approving a plan, reviewing a diff, answering a permission prompt. This subject is about the mode where nobody is watching: a script that pipes a build log into Claude and captures the answer, a CI job that responds to @claude in a PR comment, a nightly process that triages new issues while everyone's asleep. Removing the human from the loop doesn't remove the need for judgment — it just moves all of that judgment from "in the moment" to "encoded up front," in flags, workflow YAML, and guardrails that have to be right before the run starts, because there's no one there to catch a mistake as it happens.

This is squarely advanced material because it combines everything earlier subjects covered — the agentic loop, permission modes, git workflows — with a new constraint: correctness has to be verified by exit codes and JSON fields instead of a person's eyes. The subject moves through three layers that build on each other: headless mode, running claude -p non-interactively from a terminal or script; CI automation, wiring that same headless invocation into GitHub Actions or GitLab CI/CD so it runs on repository events instead of a person typing a command; and the Agent SDK, the Python/TypeScript library that exposes the identical agent loop programmatically, for when a shell invocation isn't enough control.

One boundary worth stating up front: the git mechanics this subject's automated runs rely on — commit conventions, PR creation, /code-review — are covered in the previous subject, Git Workflows with Claude Code. Here, git operations are treated as something an unattended run performs, not re-explained from scratch.