Build Log 03 — I built my AI agents a code review team

Overview

An AI-driven quality control pipeline mimics a traditional engineering organization. Rather than relying on a single model to both create and review work, a multi-agent system runs where different Claude instances perform specialized roles.

The core pipeline

The workflow consists of six stages:

1. Planner — creates implementation specifications for approval
2. Engineer — builds the feature or fix
3. Code reviewer — reviews diffs using a deliberately different model, with no access to the engineer's reasoning
4. Design & brand critics — evaluate user-facing elements
5. QA agent — attempts to disprove acceptance criteria
6. Manual approval — author decides whether to merge

What worked

The separation of reviewer from author proved most valuable. A second agent caught both a fabricated statistic and an off-brand claim the original author wouldn't flag, with no attachment to the first draft and permission to disagree until proven wrong.

What failed

The pipeline passed a landing page with visible bugs — broken navigation and dead links. The QA agent confirmed links worked by reading the href attribute rather than actually clicking them, revealing that an attribute existing is not proof a behavior works.

The critical lesson

Rather than simply fixing bugs, lessons get institutionalized as permanent rules. Three rules emerged from the navigation failure, so similar oversights don't recur through different mechanisms.

Takeaways

• Force genuinely separate review passes without visibility into prior reasoning
• Document rules, not just fixes
• Distinguish between "code correctness" and "behavioral functionality"
• Accept higher token costs as an investment in early-stage quality