The AGENTS.md also orders the agents to create relevant tests with each added code, so the tests are already present to catch regressions. The benchmarks themselves also serve as de facto tests since they'll typically break if the refactor is done incorrectly.