Crux Labs raises $6M to turn expert judgment into AI agent evaluation

The gap in agent behavior

AI agents are increasingly capable of completing complex tasks, but they often struggle with the nuances of professional judgment. An agent might successfully finish a workflow while still making a decision that an experienced employee would immediately reject. This happens because agents frequently lack access to the unwritten rules of a specific business.

These rules include knowing when to ask a clarifying question, when to escalate a problem to a human, or what a high-quality outcome actually looks like for a specific client. While engineers can build the underlying architecture, they often lack the deep domain knowledge required to define these subtle boundaries. This creates a bottleneck for companies trying to move AI from simple chatbots to autonomous agents.

Crux Labs is positioning itself to solve this by creating a bridge between human experts and engineering teams. The startup argues that businesses need evaluation systems that are specifically calibrated to their own internal standards. By capturing the heuristics held in the heads of human workers, the company hopes to create more reliable and business-aligned AI behavior.

Moving beyond simple traces

Many companies currently use agent observability to inspect what a system did during an interaction. This involves looking at the sequence of messages, tool calls, and steps taken behind the scenes. While these traces are useful for diagnosing errors after they occur, they do not provide a way to judge the quality of the interaction itself.

Crux points out that manual review of these traces cannot keep pace with large-scale production deployments. The company notes that teams often sample only 1% of agent traces, which means the vast majority of interactions go uninspected. This sampling method is particularly risky because the most important failures are often the unusual exceptions that experts recognize instantly.

The challenge lies in converting that human recognition into a repeatable, scalable evaluation. Crux wants to move away from scattered support tickets and individual reviews toward a systematic process. They aim to turn expert judgment into a continuous feedback loop that connects real-world production behavior directly to changes in the agent's performance.

How the platform calibrates

The Crux Labs workflow begins by integrating with the tools and processes a company is already using. Their agents capture expert judgments from these existing workflows to calibrate custom evaluations. These evaluations are then used to monitor production traffic and trigger alerts based on specific business requirements.

This calibration step is intended to prevent the system from rewarding a fluent but incorrect answer. A generic evaluation might give a high score to an agent that sounds confident but misses a specific organizational requirement. An evaluation grounded in expert judgment is designed to recognize the difference between an answer that sounds right and one that actually meets expectations.

As new failure patterns emerge, the platform allows organizations to incorporate that knowledge into how they assess future behavior. This ensures that the evaluation process is constantly changing and reflecting what experts learn from real interactions. The goal is to distill human judgment into a repeatable part of how the system is measured and improved over time.

Enterprise-scale deployment

Crux Labs is targeting high-growth businesses and Fortune 500 companies in complex industries such as healthcare and financial services. The company's founding team brings significant experience in deploying AI across various sectors. Co-founder Julie Kheyfets has a decade of experience in complex industries, while CTO Krishna Kulkarni has worked on scaling AI products at Meta and Checkr.

The commercial argument for the platform is that it functions as a dedicated agent evaluation and improvement team for the enterprise. By using Crux, product and engineering staff can spend less time diagnosing recurring failures and assembling review processes. Meanwhile, subject matter experts gain a practical way to influence the systems that represent their business operations.

The company plans to use its new funding to expand its research and engineering capabilities while accelerating deployments. While the company has not disclosed specific revenue or customer names, it reports early traction in several complex fields. The ultimate goal is to ensure that the judgment of a company's best people becomes a core part of how its agents behave in production.