Design review rather than code review

A development head at a business software company has a person review an AI agent's plan before code exists, so code review checks mainly for unrequested changes.

The head of development at a business software company, writing as tanashun, published an account in Japanese of how the team reviews code from Claude Code, Anthropic's AI coding agent. Pull requests were arriving faster than anyone could read them, and design mistakes found late forced large rewrites. The team split human review into a design review of the AI agent's plan and a narrower code review.

• Before coding, a person reads the AI agent's plan and checks its fit with the requirements, the code it touches, and the shape of data and interfaces.
• The approved plan becomes a work order, and the AI agent that writes the code may not change the design.
• AI reviews check that the code follows the plan: a light review after each change, and a heavier review in parallel before the pull request.
• In code review, the person looks mainly for anything the AI agent added that the requirements did not ask for, and the author says the extra effort is worth it.

The author argues that a team which cannot read every AI-written pull request should put the human reading into a short plan, where a wrong direction costs less to fix. The account reports no measurements, and the author also mentions a personal app and offers consulting work.

3 November 2026: teams working with AI agents are moving human review from each output to the plan, the shared instructions and the rules the agent follows beforehand, and naming in advance who answers for the result.