Agents in the pipeline, in brief
Lesson 7 takes the human out of the loop. Everything the track has built so far, configuration, schemas, tool contracts, orchestration, reliability, gets its sternest test in a setting where nobody is at the keyboard: Claude Code running headless inside a CI pipeline.
The capability: after this lesson, you can design an automated pipeline job around Claude Code, a nightly batch task or a blocking pre-merge check, and defend its permissions, context, output format, and budget.
What the lesson covers. Claude Code’s non-interactive mode runs one prompt and exits (the print flag), reading standard input and writing standard output like any command-line tool. A pipeline run can never hang waiting for input, so permissions are granted in advance. Output format flags return structured JSON a script can parse and post as pull request comments, with a schema option that fixes the shape of findings. The CI agent inherits the checked-in configuration layer from lesson 2, project memory and rules files, which makes it the newest teammate rather than a special case. The heart of the lesson is the generator-reviewer split: an instance reviewing code it just generated carries the reasoning that produced the mistakes, so the reviewer must be a separate invocation with fresh context. That is lesson 6’s independence principle turned into mechanics. Around it: deduplicating findings on re-runs by feeding prior comments back into context, writing review criteria as concrete examples rather than moods (because a false-positive flood destroys developer trust), and giving batch jobs and blocking checks opposite cost and latency budgets.
Why this order. The track spine says architecture is deciding where judgment lives. An unattended agent is the purest test of those decisions: with nobody there to catch it, only the checked-in design is left. Lesson 8, the capstone, asks you to make those decisions on a system of your own and defend them.