Make two agents plan, then trade their homework

I don’t trust a single agent’s first plan, whether it came from the model or from me. So for anything past a trivial change, I make two of them do the thinking before any code exists.

I describe the task to Cursor’s Composer and to Claude Code separately, and let each write its own implementation plan. Then I swap them. Composer reviews Claude’s plan, Claude reviews Composer’s. Each one reliably catches something the other took for granted. A database migration nobody mentioned. An edge case in the auth flow. A wrong assumption about where state actually lives.

The first time I ran this properly was on a feature that touched billing, which is exactly the kind of thing you never want an agent freelancing on. Composer’s plan looked clean until Claude pointed out it would double-charge anyone who retried a failed payment. That’s a bug I would have shipped and then heard about from a very annoyed customer at the worst possible hour.

The merged plan is consistently better than either draft, and the whole thing costs me about five minutes and two prompts. It works because the two models fail in different places. Two reviewers who make different mistakes catch more between them than one confident reviewer making the same mistake twice.