A conversation with a single AI agent can seem flawless in isolation, yet it might still indicate underlying product issues. This discrepancy is pushing enterprises to shift from evaluating individual interactions to analyzing user cohorts against a performance baseline. At the VB Transform 2026 event, leaders from LangChain, Conviva, and CoreWeave discussed this shift, highlighting a move towards using more cost-effective, purpose-specific judge models to evaluate AI agents. Despite advances in automated judging by large language models (LLMs) or agents, human review remains crucial for accuracy and trustworthiness. Harrison Chase of LangChain emphasized that evaluation criteria should act as evolving product specifications rather than fixed tests, fostering continuous improvement. Hui Zhang from Conviva explained that scoring conversations one at a time misses patterns only visible by comparing groups, a method known as contrastive analysis. Emmanuel Turlay from CoreWeave noted that widespread and ongoing monitoring detects real-world failures more effectively than pre-launch testing alone. Furthermore, scalable models tuned to detect errors—such as LangChain’s fine-tuned Qwen model—offer significant efficiency gains. However, human oversight is still essential for accountability, especially in critical sectors like legal, healthcare, and finance. The panel highlighted the importance of human involvement not only for safety but also for building trust and enabling AI systems to learn over time.
Back
