Introduction

In the previous lesson, you integrated a single agentic code-review step into a continuous-integration pipeline using GitHub Copilot. That exercise established the four architectural pillars: event trigger, authentication, context assembly, and result publication, and then demonstrated that an AI agent can operate as an ordinary CI job on an iOS pull request.

A production team rarely standardizes on one provider indefinitely. Model behavior, token consumption, latency, and the quality of iOS-specific reasoning vary across vendors. The feedback produced by any agent is probabilistic; the same diff can elicit different findings from different models or even successive runs of the same model. Consequently, the engineering organization must develop both the technical ability to invoke multiple agents, and have the judgment required to interpret, prioritize, and act on their output.

This lesson expands the pipeline to include two additional widely available agents:

• Claude Code (via the official Anthropic GitHub Action)
• OpenAI Codex (via the Codex GitHub Action)

These are new tools for this program. They are not the Claude Agent in Xcode or Cursor from earlier modules. If you have not already created the Anthropic and OpenAI accounts and repository secrets described in the previous lesson, do that before this demonstration.

You’ll submit the identical pull-request diff to each of them, and review a side-by-side comparison of the resulting comments, token consumption, and wall-clock latency. This reveals the practical differences that cannot be predicted from marketing claims alone. Finally, you’ll understand how a human reviewer should triage AI-generated comments, what constitutes a useful versus a noisy finding, find under which conditions agentic review reliably improves outcomes, and when it adds token usage without proportional value.

By the end of this lesson you’ll be able to:

• Configure Claude Code and Codex as alternative or complementary CI review agents using the same architectural patterns established previously.
• Execute the identical iOS pull-request diff against multiple agents and capture comparable metrics for quality, latency, and token count.
• Formulate the concrete criteria necessary for comparing agent output on Swift concurrency, memory management, SwiftUI state ownership, and project-specific goals.
• Apply a practical triage policy that classifies findings as:
  • actionable defects,
  • discussion points,
  • or dismissible noise.
• Identify classes of change where agentic review is most and least effective:
  • small focused fixes,
  • large architectural refactors,
  • security sensitive paths,
  • documentation-only updates.
• Recognize the limits of automated review and the continuing necessity of human judgment for merge decisions.

The written instruction develops the comparative framework and evaluation principles in a way that can be applied as AIs evolve. The accompanying demonstration will apply those principles to a concrete iOS pull request, recording the differences that appear in practice.