The Two-Claude Relay

A working pattern for an autonomous build loop: one agent executes, a different agent reviews, and neither can ship alone.

This was built and run in production. It also took a site down for four hours, deadlocked twice, and came one click away from git reset --hard on the only copy of a finished feature. The safety rules in autonomous-loop-safety are not optional garnish — they are the difference between this working and this costing you a night. Read that skill alongside this one.

The core idea

One agent does the work. A different agent checks it. They communicate through two files in a git repo. A shell script keeps the loop turning.

The split matters more than it sounds. An agent verifying its own work grades its own homework — it knows what it meant to do, so it tends to see what it meant to do. A separate reviewer, with a separate context window and no memory of writing the code, has no such loyalty.

That isn't theoretical. On the first run, the executor's overflow test came back clean — but the entire UI it was testing lived inside a display:none section, so the probe was measuring hidden elements and reporting a beautiful green pass on exactly the components under test. A single agent ships that false green.

The division of authority

This is the safety model, not ceremony: