From Dev.to
Terminal-Bench: Why Agent Benchmarks Need Continuous Integration to Stay Valid
Coding agent benchmarks rot. Not because the tasks change, but because models discover valid solutions that verifiers reject. A benchmark carefully validated at release can become stale within months...