Crux Labs 融資逾 600 萬美元，將企業專家判斷轉化為更完善的 AI 智能體

An AI agent can complete a task and still make a decision an experienced employee would reject. The difficult part is often understanding the unwritten rules: when to ask another question, when to escalate, and what a good outcome means for a particular business.

Crux Labs is building a platform to turn that expert judgment into a continuous evaluation and improvement process for enterprise agents. The New York startup has emerged from stealth with more than $6 million in an oversubscribed funding round, co-led by Benchstrength and Misneach, with participation from Analog Ventures and Pareto Holdings.

The company plans to use the capital to expand research and engineering and accelerate enterprise deployments. Its central bet is that businesses need evaluation systems calibrated to their own experts, with a feedback loop that connects production behavior to changes in the agents themselves.

Why Agent Traces Are Only the Starting Point

Agent observability helps teams inspect what a system did: the sequence of messages, tool calls, and other steps behind an interaction. That record is valuable when something goes wrong. Deciding whether the interaction was good enough, however, requires criteria that reflect the task and the business.

Crux argues that manual review cannot keep pace with large deployments. Its announcement describes teams sampling just 1% of agent traces, leaving much of the remaining traffic outside direct inspection. That figure is the company’s description of the problem, rather than a published industry-wide survey result.

Sampling can be especially limiting when the failures that matter are unusual. An agent may behave acceptably in common cases while struggling with an exception that experienced staff recognize immediately. The challenge is converting that recognition into an evaluation that can be applied consistently across a much larger set of interactions.

For Crux, the opportunity lies between the people who know the domain and the engineers who build the agent. Subject matter experts can explain what a successful interaction looks like, but that understanding needs to become a repeatable part of how the system is measured and improved.

Capturing Judgment and Calibrating Evaluations

Crux describes a workflow that begins inside the tools and processes a company already uses. Its agents capture expert judgments and use them to calibrate custom evaluations. Those evaluations then support production monitoring and alerts, with the results feeding subsequent agent improvements.

The calibration step is important. A generic evaluation might reward a fluent answer that misses a business-specific requirement. An evaluation grounded in expert judgment is intended to recognize the distinction between an answer that sounds right and one that meets the organization’s expectations.

“Agents are incredible at attaining goals, but they lack access to the years of heuristics that we as humans have collected in our heads,” said Julie Kheyfets, Crux’s co-founder and CEO. “Crux helps every enterprise to distill this judgment and to use it to rapidly improve their agents.”

The intended outcome is a changing evaluation process that reflects what experts learn from real interactions. As new failure patterns emerge, the organization has a way to incorporate that knowledge into how it assesses future behavior, rather than leaving it scattered across support tickets and individual reviewers.

Monitoring the Full Stream of Production Traffic

Crux says its custom evaluations are optimized for real-time monitoring and alerting across 100% of production traffic. That is a claim about coverage: applying evaluation across the traffic stream. It does not mean the evaluations detect every possible failure or agree perfectly with human experts.

The distinction matters because automated evaluation introduces its own measurement challenge. A system that reviews every interaction can still miss important behavior if its criteria are incomplete. Calibrating those criteria against domain expertise is therefore central to Crux’s proposition, rather than an optional step before monitoring begins.

The company also says evaluation results drive improvements shipped into production. Its announcement does not detail which parts of that process change prompts, policies, models, or other components, or how customer approval is handled. The disclosed proposition is an end-to-end evaluation and improvement service, rather than a specified model-training technique.

A Founding Team With Enterprise AI Experience

Kheyfets and co-founder and CTO Krishna Kulkarni met at Princeton University. Kheyfets brings a decade of experience deploying AI in complex industries, including leading North America for Tractable AI and serving as CEO of Block Renovation.

Kulkarni previously led brand engineering at Meta, developed AI products at Checkr, and helped scale Healthify from zero to 20 million patients, according to the company. Crux is building a team of researchers and engineers focused on evaluating and improving agents.

“Engineering teams know how to build AI agents, and subject matter experts know what success looks like, but enterprises lack the bridge between them,” Kulkarni said. “Crux operationalizes domain expertise into continuous agent improvement so engineers can build truly business-aligned agents.”

The company reports early traction with Fortune 500 enterprises and high-growth businesses in healthcare, cybersecurity, financial services, and other complex industries. It has not named those customers or disclosed revenue, deployment counts, or measured improvements in the announcement.

The Business Case for Company-Specific Evaluation

Crux is positioning itself as the equivalent of a dedicated agent evaluation and improvement team inside an enterprise. The commercial argument is that product and engineering staff should spend less time assembling review processes and diagnosing recurring failures, while experts have a practical way to influence the systems representing their business.

The demand for domain expertise also appears in the iMerit platform for expert-led AI training and evaluation, where organizing specialist feedback is part of the model-improvement workflow.

For investors, the opportunity is tied to what happens after agents move into production. Building an agent creates a new need to measure its behavior continuously, identify where it falls short, and improve it without losing sight of the organization’s standards.

Crux’s funding will help test whether that process can be delivered reliably across customers with different definitions of success. Its proposition is straightforward: the judgment held by a company’s best people should become part of how its agents are evaluated, and those evaluations should lead to better behavior in production. More information is available at Crux Labs.