Beyond the "Vibe Check": The Indispensable Role of Evals in AI's Next Frontier

The assertion that a multi-million dollar AI coding agent business was built largely on "vibes" rather than rigorous evaluations ignited a fervent debate among AI professionals. This discussion, hosted by Swyx on the Latent Space podcast, brought together Ankur Goyal, co-founder and CEO of Braintrust, and Malte Ubl, CTO of Vercel, to dissect the critical role of evaluations (evals) in the rapidly evolving landscape of AI engineering. Their conversation transcended a simple dichotomy, revealing a nuanced spectrum of feedback loops, from intuitive "vibe checks" to complex offline evaluations and A/B testing, each playing a distinct, deliberate role in driving AI product development.

At its core, the challenge of building AI products lies in grappling with "non-deterministic magic," as Ankur Goyal aptly puts it. Unlike traditional software development, where outcomes are often predictable, AI introduces an inherent uncertainty that demands robust feedback mechanisms. The panelists emphasized that the goal isn't to choose one feedback loop over another, but to strategically deploy a combination of approaches, leveraging their unique trade-offs in effort, speed, and efficiency.

Malte Ubl, drawing from his experience at Google Search and Vercel, highlighted the foundational importance of understanding whether a change constitutes an improvement. "I want to know if I'm doing well, and how fast can I find out," he stated, encapsulating the developer's quest for immediate, actionable feedback. This need for velocity is precisely where a diverse set of feedback loops becomes indispensable. While a "vibe check" offers instant, albeit subjective, insight for small, rapid iterations, more structured methods become necessary for scaling and ensuring long-term product health.

Offline evaluations, once centered around static "golden datasets," have evolved significantly. Top teams now dynamically pull real user failures from production logs, transforming them into daily additions to their eval suites. This iterative process allows for continuous learning and adaptation, ensuring that evaluations remain relevant to actual user experience rather than becoming stale or misaligned. Such a production-driven approach not only enables faster iteration but also instills confidence, allowing teams to ship aggressively without fear of regression.

Coding agents present a uniquely verifiable use case for evaluations. Signals like "does it compile?" or "does it render without errors?" provide objective, binary outcomes that are highly amenable to automated testing. Vercel, for instance, leverages these verifiable signals in Reinforcement Learning (RL) pipelines to fine-tune models that can fix trivial errors orders of magnitude faster than human intervention or complex agentic loops. This automation of error correction is a powerful demonstration of how well-designed evals can dramatically accelerate development cycles and improve product quality.

The discussion also illuminated the emerging role of evals as a product management tool. Rather than relying solely on lengthy Product Requirement Documents (PRDs), product managers are now deeply involved in designing evaluation rubrics and LLM-as-judge scoring functions. This allows them to encode domain expertise, whether in finance, healthcare, or other specialized fields, with greater precision and directness, translating nuanced product requirements into measurable AI performance. This shift empowers PMs to communicate desired outcomes more effectively and ensures alignment between business goals and AI development.

A critical insight from the conversation touched upon the "privilege of AI labs." Organizations like Anthropic, with their vast resources and integrated research capabilities, can build proprietary, in-house evaluation systems that serve as significant competitive moats. These internal benchmarks are often far more trusted than public benchmarks, which, while useful for marketing and general comparison, rarely capture the specific nuances and performance metrics crucial for a company's unique product. This disparity underscores the strategic advantage held by large AI labs in defining and measuring their own success.

Looking ahead, the panelists pointed to RL environments as the next frontier for evaluations, particularly for computer-use agents. These environments promise to decouple evaluation from expensive human labeling, offering a powerful, scalable method for assessing agent performance. However, they also demand specialized expertise to avoid "reward hacking," where an agent optimizes for the metric rather than the true underlying goal. An intriguing idea for the future is an "inversion of control" for evals, where platform companies like Vercel publish Next.js evals, allowing model labs to optimize their AI agents specifically for those frameworks, fostering a new marketplace where eval creators and model trainers are distinct entities.

Ultimately, the debate concluded not by dismissing "vibes" but by integrating them into a comprehensive framework of feedback loops. Vibes, while subjective, are acknowledged as an extraordinarily accurate, albeit expensive, scoring function. The most successful teams, therefore, are those that deliberately invest in multiple feedback loops, from rapid vibe checks to statistically significant A/B tests and robust offline evaluations, calibrating their investment to the specific needs and maturity of their AI products. This holistic approach ensures continuous improvement, rapid iteration, and a deeper understanding of AI performance in the real world.