At the VB Transform 2026 conference in July 2026, leaders from LangChain, Conviva, and CoreWeave pointed out a major paradox in AI development: a single AI agent conversation can look flawless in isolation, yet the underlying product remains broken. This gap is driving a significant shift in how enterprises evaluate AI agents, moving away from scoring individual traces and toward comparing cohorts of users against a baseline.
Background & Causes
Falling into "eval paralysis" before releasing a product is a common pitfall for development teams. According to Harrison Chase, CEO of LangChain, many teams attempt to build an exhaustive evaluation suite and refuse to launch until it is perfect, whereas the best teams launch first and iterate based on reality. Similarly, Emmanuel Turlay from CoreWeave shared his experience of trying to achieve 100% test coverage only to still encounter serious bugs in production.
Technical & Technology Analysis
The core weakness of traditional evaluation methods lies in analyzing traces in isolation. Hui Zhang, CTO of Conviva, advocated for "contrastive analysis" by comparing user cohorts against a baseline. For instance, a shopper successfully purchasing running shoes after answering qualifying questions might seem like a perfect interaction; however, if the clarification ratio for this cohort is three times higher than the baseline, there is a systemic issue that individual scoring misses. Furthermore, fine-tuning smaller open-source models, such as Alibaba's Qwen, through distillation has allowed LangChain to achieve a 10x to 100x cost reduction compared to using heavyweights like Claude Sonnet as judges.
Expert Opinions & Insights
The debate between automated evaluation (LLM-as-judge or Agent-as-judge) and human review remains unresolved. Zhang noted, "You have scalable but ungrounded, whether it's agents as judge or LLMs as judge... the whole industry is facing this, which poison you want to pick." Meanwhile, keeping humans in the loop remains critical, especially in high-stakes fields like law, finance, and healthcare. Turlay emphasized that before we can completely remove humans from the sign-off process and legal accountability, the industry still has a long way to go.
Impact & Future
The shift toward smaller, specialized, and cost-effective judge models is reshaping how AI applications are built. Developers should recognize that evaluation criteria (evals) now function as a living product requirements document (PRD) that evolves alongside real-world user feedback. Combining broad online monitoring with human oversight on corner cases will be key to building trustworthy AI agent systems in the future.