AI agent evaluation shifts from clean traces to cohort baselines

Enterprises evaluating AI agents are moving away from grading single conversations and toward comparing groups of users against baseline behavior, executives from LangChain, Conviva and CoreWeave said at VB Transform 2026. The shift reflects a practical problem for agent teams: an individual transcript can look successful while the product is still failing in a measurable way.

Harrison Chase, CEO of LangChain; Hui Zhang, CTO and co-founder of Conviva; and Emmanuel Turlay, director of engineering at CoreWeave, also described a related cost shift. Teams are still using large language models as judges, but they are starting to route narrower review tasks to smaller, cheaper models or even non-ML rules when the job allows it.

Evaluation is becoming a working product spec

Chase said LLM-as-judge remains the common default, even as some teams experiment with agent-as-judge setups. Zhang framed the harder tradeoff as automated review versus human review: automated systems scale but can be poorly grounded, while humans provide judgment that is expensive and difficult to apply broadly.

For product teams, Chase argued, evaluation criteria increasingly serve the role that product requirements documents have played in conventional software development. They describe what the agent is supposed to do, what it should refuse and where behavior needs to change after launch.

He warned against teams waiting to build a complete evaluation suite before shipping. In his view, stronger teams release, observe failures and revise their checks from live behavior. Turlay described a similar software testing problem from his own experience: even test suites built for full coverage still missed production bugs.

Turlay said broad online monitoring should come before a narrow offline benchmark. That lets teams spot actual failure categories in use, then build targeted evaluations around those recurring problems rather than trying to predict every edge case ahead of time.

Single traces can hide product failures

Zhang criticized the common method of sampling individual traces and scoring them one by one, whether the sample is small or comprehensive. Conviva’s preferred approach, which Zhang called contrastive analysis, compares user cohorts against a baseline to expose patterns that do not appear in an isolated conversation.

He gave a retail example involving a shopper asking an agent for running shoes before a half marathon. The agent asks follow-up questions, the shopper buys a shoe and the conversation appears successful when reviewed alone.

Across the broader user population, however, Zhang said the agent asked three times as many clarifying questions as the baseline for that shoe category. He also said shoppers completed purchases outside the conversation at five times the baseline rate for the same category. Those metrics point to a category-level defect that a clean transcript would miss.

Zhang added that many agent evaluations lack context from before, between and after conversations, relying too heavily on the transcript itself.

Judge models are being sized down

Turlay said teams should begin with the strongest available model to test whether a judgment task is feasible, then move down to cheaper options. If a task cannot be handled by a top model, he said, a smaller one is unlikely to solve it. Once the pattern is proven, teams can judge a sample of traffic and send simpler tasks, such as binary classification, to smaller open source models.

LangChain has followed that path. Chase said the company fine-tuned a Qwen model from Alibaba’s open source family to detect “perceived error,” meaning cases where a user appears to believe the agent made a mistake. Using human labels and distillation, Chase said the model matched Claude Sonnet on the task while cutting serving costs by 10 to 100 times, depending on deployment.

Chase also said some guardrails do not require a model at all. He pointed to Claude Code using regex-based checks for some safeguards, a reminder that deterministic rules still have a role when the pattern is clear.

Humans remain in the approval chain

The speakers did not argue that model-based judging removes human responsibility. Turlay cited his previous work in autonomous vehicles, where his team reduced data intake and retraining into a two-week cycle before shipping a model to a car. Even with that process, a person still had to approve release.

He said regulated or high-liability fields such as legal, finance and healthcare will keep requiring someone to accept responsibility before agents act on their own. Zhang said humans remain necessary for corner cases, while automation can detect patterns at a scale individual reviewers cannot match. Chase added that human review also helps build trust and provides feedback that agent systems can learn from over time.

This story draws on original reporting from VentureBeat.