Advanced AI Eval Techniques: Prioritizing Error Discovery Over Metric-First Approaches

AI product teams are increasingly asked to write “evals” – repeatable tests that capture what “good” looks like for a model. Yet the same source warns that most teams skip the first, arguably most crucial, stage: error discovery.

The hidden cost of metric-first thinking

In the original “Building eval systems” post, the authors described a three-stage process: discover errors, create metrics, and close the loop. The new article observes that teams routinely bypass the discovery phase, opting instead to define metrics from the outset. This shortcut appears attractive because metrics are concrete and can be automated, while sifting through user session traces feels labor-intensive. However, the source argues that early metric design forces teams to make assumptions about what matters, often resulting in “measuring the wrong thing, or the right thing poorly.”

Error discovery is framed as the eval equivalent of product discovery: it surfaces the failures that truly affect user outcomes. Without it, dashboards become collections of generic numbers that waste engineering time and steer development toward irrelevant goals. The author stresses that if a team can only afford one part of the eval process, error discovery should be that part.

A 30-minute, three-step workflow with coding agents

The newsletter outlines a repeatable, 30-minute workflow that leverages a coding agent (e.g., Codex or Claude) and an “eval skills” plugin. The process has been trialed with more than 50 AI companies and consistently uncovered “major product flaws that were hurting the customer experience.”

1. Start with diverse traces – A trace bundles the user input, system prompt, intermediate calls, and final output. The guide recommends sampling a varied set of traces to capture the full behavior spectrum.
2. Active-learning loop – Review the sampled traces manually, flagging any failure. Then examine a few more instances of the same failure to confirm its pattern before asking the agent to annotate the rest. This balances human judgment with the speed of automation.
3. Agent-assisted annotation – Once the failure type is defined, the coding agent can label additional traces, allowing the reviewer to accept or reject each suggestion. The plugin automates bookkeeping, sampling, and annotation, keeping the loop efficient.

The source emphasizes that agents excel at catching “obvious” errors—e.g., contradictions between model output and tool results—while they miss failures that require product-specific judgment, such as “objection handling” in an AI leasing assistant. Human review therefore remains essential to prevent “criteria drift,” where the definition of a good response evolves after seeing new examples.

Real-world impact across leading AI products

The article cites several companies that have leveraged eval-driven error discovery to achieve measurable gains:

• Cursor improved user satisfaction and reduced costs by 41 % after applying evals to its Auto Balance routing system.
• Ramp lifted on-device receipt-matching precision from 35 % to 83 % through eval-guided model refinements.
• Harvey nearly doubled its internal quality score after rebuilding its AI contract reviewer with evals.

Additional firms—including Rippling, Glean, Abridge, ElevenLabs, and Robinhood—have reported systematic product improvements via evals, underscoring the broader industry momentum.

Balancing automation with human insight

A broader study mentioned in the source evaluated 100 production traces from an AI leasing assistant. Agents missed issues that required contextual judgment (e.g., Markdown formatting, human handoffs) but excelled at flagging obvious contradictions. They also introduced noise by labeling correct responses as failures. The authors conclude that the optimal approach pairs automated agents with a human-in-the-loop review, guided by active learning to focus effort on the most informative examples.

The workflow’s reliance on a coding agent does not eliminate manual work; instead, it restructures it. Humans define success criteria, validate a small sample, and then delegate bulk annotation to the agent. This hybrid model reduces the time spent on repetitive labeling while preserving the nuanced understanding needed for product-specific failures.

In sum, the newsletter reframes error discovery from a peripheral task to the cornerstone of any AI eval strategy. By allocating a brief, structured session to trace analysis and leveraging coding agents for scale, product teams can surface high-impact bugs early, align metrics with real user outcomes, and ultimately ship more reliable AI features.