LangSmith cuts agent-evaluator cost 82%

LangChain launched LangSmith Tuned Evaluators for production traces, reporting that its Perceived Error judge beats every frontier model in its benchmark while cutting evaluation cost by 82%; some early partner workloads saved 98%. The product announcement says the evaluator becomes eligible after two human–AI message pairs and completes within 12 hours, so the gain is broader coverage rather than instant blocking. Teams measuring agent quality should price evaluator coverage per eligible conversation, extending the archive’s case for regression loops around agent behavior.