Introduction

In the first two lessons of this module, you established agentic code review as a first class continuous-integration step, configured multiple providers against the same iOS pull_request diff, and practiced a disciplined triage policy that classifies findings as confirmed defects, discussion points or dismissible noise. Those capabilities are necessary but not yet sufficient for sustained use inside a real engineering organization.

Once automated reviews appear on every pull request, new operational questions surface immediately:

• How do you prevent the cumulative token count from growing without bound as the team scales?
• How do you ensure that the review job coexists cleanly with existing build, rather than competing for runner capacity or polluting the conversation with redundant comments?
• How do you turn the pattern of triage decisions into systematic improvement of the prompts, guidance files and path filters so that the signal-to-noise ratio doesn’t become static?

This lesson addresses the above questions. You’ll examine practical governance controls—path filters, draft-pull-request exemptions, concurrency and timeout policies, token-aware model selection—and you’ll place the agentic review job inside a realistic multi-job GitHub Actions workflow that already builds and tests the starter project. You’ll then close the feedback loop: recording triage outcomes, refining the shared CODE_REVIEW.md (or equivalent guidance), and adjusting prompts so that future runs of the same agents produce fewer dismissible comments and more actionable ones.

By the end of this lesson, you’ll:

• Design path filters, concurrency groups and draft pull request exemptions that keep token consumption proportional to risk.
• Integrate an agentic review job into a multi-job iOS CI workflow without introducing duplicate comments.
• Apply least-privilege permissions and secret-handling practices appropriate for production runners.
• Show triage outcomes and translate them into concrete updates to repository guidance files and agent prompts.
• Recognize the organizational signals (rising false-positive rates, latency spikes, token-count anomalies) that indicate the need to re-calibrate or temporarily disable an agent.
• Establish a lightweight frequency for reviewing the effectiveness of the automated review channel itself.

The written instruction develops the governance and feedback-loop principles in evergreen form, just like in the previous lessons. The demonstration then applies those principles to the sample iOS repository, extending the workflows created in the previous lessons and showing a concrete refinement cycle.

Ready?