The 3-Agent Setup for Code Review: One Writes, One Reviews, One Watches

One AI agent writing code, another reviewing, a third watching for drift. Here is when the 3-agent setup earns its overhead and when it doesn't.

The 3 roles

Writer

Standard Claude Code session. Given a task, produces the diff. Nothing unusual.

Adversarial reviewer

Same task file, different session. Prompt: "Find 3 things wrong with the diff in staged changes. Be specific." Does NOT know it's reviewing an AI's output; treats it as any code review.

Drift-watcher

Third session, running periodically (every 20 minutes). Prompt: "Check if the changes to match the original task description. Flag if scope drifted."

Why 3, not 2

Two agents (writer + reviewer) tend to converge: the reviewer catches syntax and obvious bugs, but not scope creep. The drift-watcher, with distance from the diff, catches "we started fixing bug X and ended up rewriting module Y".

When it pays off

• Multi-hour tasks where you'll walk away
• Refactors with fuzzy scope
• Anything that touches auth, payments, data migration
• When you're too tired to review carefully yourself

When it's overkill

• Single-file bug fixes
• Adding tests to existing code
• Tasks under 30 minutes

The overhead cost

• 3× token cost
• ~3× wall time to keep all in sync
• Attention cost (you review reviews of reviews)

Making it manageable

You need a tiled setup where each session is visible at a glance — otherwise you'll forget the reviewer or the drift-watcher exists. Named panels, color per role, ⌘-click broadcast to sync task descriptions.

Agent Desk was built for exactly this pattern: named tiled panels, per-role color, crash-proof restore. $10 once, full source ships with it.