A developer watches a single trace of their AI agent and sees a flawless execution. The agent understands the intent, calls the correct tool, and delivers a polished response. On paper, the test is a success. Yet, when this same agent hits production, the user experience collapses. This gap between a perfect individual conversation and a failing product is the primary friction point currently stalling the deployment of autonomous agents across the enterprise.
The Shift Toward Contrastive Analysis
At VB Transform 2026, a candid discussion between Harrison Chase, CEO of LangChain, Hui Zhang, CTO of Conviva, and Emmanuel Tourle, Engineering Director at CoreWeave, highlighted a fundamental flaw in how the industry measures AI success. For too long, the gold standard has been the individual trace—scoring a single interaction on a scale of correctness. However, Zhang argues that this granular focus creates a blind spot. He proposes a transition to contrastive analysis, where the unit of measurement is the cohort rather than the conversation.
Consider a retail agent designed to help customers buy shoes. In a single-trace evaluation, an agent that successfully leads a user to a purchase receives a high score. But when viewed through a cohort lens, a different story emerges. If the data shows that this agent asks three times as many confirmation questions as the baseline, or if users are completing their purchases outside the chat interface five times more often than they used to, the product is failing despite the high individual scores. The friction is invisible in a single trace but glaring in a population sample.
This shift in perspective changes the development lifecycle. Harrison Chase suggests that evaluation sets should no longer be treated as a final exam that must be passed before release. Instead, they should function as living specifications, similar to a Product Requirement Document (PRD). By treating evals as evolving documents, teams can avoid evaluation paralysis—the state of delaying a launch indefinitely in pursuit of a perfect test set. The strategy is to ship early and use real-world data to iteratively refine the specification.
The Hierarchy of Judges and the Evaluation Paradox
As the industry moves toward these broader metrics, the cost of judging those metrics has become a bottleneck. The current trend in judge model adoption is a move from general-purpose high-performance models toward specialized, smaller, and cheaper architectures. Emmanuel Tourle advocates for a hierarchical approach: prove the feasibility of a task using the most powerful model available, then systematically migrate the task down to the smallest model capable of maintaining that performance. If a problem cannot be solved by the top-tier model, it is a waste of resources to attempt it with a smaller one.
LangChain has already put this into practice by addressing perceived error—the moment a user feels the agent has made a mistake. Rather than relying on expensive frontier models, LangChain fine-tuned Qwen, an open-source model from Alibaba. By combining manual labeling with distillation techniques, they achieved performance comparable to Claude Sonnet while reducing operational costs by 10 to 100 times. This demonstrates that the judge does not need to be a generalist; it only needs to be an expert in identifying specific failure patterns.
Furthermore, the reliance on LLMs for every guardrail is being questioned. In the development of Claude Code, engineers found that traditional programming techniques often outperform models in efficiency and reliability. Specifically, they utilized regular expressions (Regex) to validate patterns and implement guardrails. This choice underscores a critical realization in AI engineering: the most efficient tool for a task is often not an LLM, but a deterministic piece of code.
This leads to what Tourle describes as the evaluation paradox. Many teams strive for 100% test coverage in their offline sets, only to find that production environments still spawn unpredictable bugs. The solution is to invert the workflow. Instead of trying to predict every failure offline, developers should build broad online monitoring systems to identify actual failures in the wild. These real-world failures then become the targeted data points used to build more robust offline evaluation sets.
Ultimately, this automated pipeline cannot exist in a vacuum. In high-stakes sectors such as law, finance, and medicine, the concept of human-in-the-loop remains non-negotiable. Because a model cannot assume legal or ethical liability, a human must remain the final arbiter of truth. Human intervention is not merely a safety net; it is the primary mechanism through which the system learns, remembers, and earns user trust. The goal for the modern AI practitioner is to design a virtuous cycle where machines identify patterns of error and humans validate them, feeding those insights back into the system's core logic.
