Tag
This paper presents a coroutine-bridge harness that reduces model calls for LLM agents on CAR-bench, enabling policy enforcement as code and achieving top performance with minimal latency.