Building & appsClaude

⚔️ Agent teams: a review crew that argues

One AI reviewing its own work has a blind spot: it agrees with itself. The fix used by serious builders: agent teams — separate Claude instances with opposing jobs, forced to argue before anything ships. Claude Code supports this directly with subagents; the pattern works anywhere. Here's the review team setup.

Why arguing agents beat one smart agent

Quality comes from tension. A builder wants to ship; a critic wants to find flaws; a user advocate wants simplicity. One prompt can't hold all three stances honestly — it averages them. Separate agents with separate incentives actually fight, and the work gets better in the crossfire.

The three-agent review team

We're using an agent team review for this project. Set up three subagents with these charters: 1. BUILDER — implements the feature. Cares about: working code, shipping today. Presents the work and defends its choices. 2. CRITIC — reviews everything the Builder makes. Cares about: what breaks, edge cases, security holes, sloppy shortcuts. Must find at least 3 real issues per review — "looks good" is a failed review. 3. USER ADVOCATE — reviews from the perspective of a non-technical person using this. Cares about: confusion, too many steps, unclear labels, what happens when someone does the unexpected. Process: Builder presents → Critic attacks → Builder responds or fixes → User Advocate reviews the result → summarize the argument and what changed. Show me the debate, not just the conclusion.

Reading the debate (your job)

The transcript of the argument is the deliverable. You'll see the Critic flag something, the Builder push back, and — this is the valuable part — the places where they can't resolve it. Those unresolved tensions are the real decisions, and they're yours to make.

Variations worth running

Design review: swap the charters for the 4 design roles and let them argue about your homepage.
Decision review: Optimist / Pessimist / Accountant arguing a business decision (thinking-partner guide).
Pre-launch: Critic and User Advocate only, on the finished thing, the day before it ships.

Agent teams multiply tokens and time. A three-agent debate costs real compute versus one prompt. Use the team for things that matter shipping wrong — not for every button label.

Force the Critic's quota. "At least 3 real issues, every review" is the line that keeps the team honest — without it, agents drift back toward agreeing with each other by round three.

Want this done for you?

We set up these exact systems for businesses like yours. One call, honest scope, no pressure.

Start your project