Building Production-Grade Agent Evals

The Breakthrough

Reliable agent behavior in non-deterministic systems is achieved by treating evaluation as a continuous feedback loop rather than a static scorecard, specifically by using agent trace logs to diagnose reasoning failures rather than just output outcomes.

What Actually Worked

• Start with intuition-based vibing: In the early stages, prioritize qualitative assessment of agent outputs to identify failure patterns before building expensive, scalable evaluation sets.
• Implement a remediation loop: Build an independent critique agent that acts as a self-correction mechanism to fill gaps in the base tool set.
• Analyze agent traces: When an agent fails a task, such as removing a mandatory legal disclaimer, inspect the internal trace logs to determine if the model correctly identified the constraint but failed the execution logic.
• Focus on patterns over isolated runs: Avoid hyper-fixating on fixing single prompt failures, which can lead to overfitting; instead, use a golden set of examples to measure the frequency of failure patterns across the entire system.
• Standardize human-in-the-loop rating: Provide human raters with clear rubrics and specific examples to ensure high inter-rater agreement, and require written explanations for ratings to understand the reasoning behind pass or fail decisions.

Context

The Google YouTube Ads team encountered significant reliability issues while building an agent designed to clean up messy advertising assets. Because generative models are non-deterministic, they found that prompt engineering alone could not guarantee compliance with strict requirements like mandatory legal disclaimers. The team shifted to a rigorous evaluation framework that combines human-led golden sets, LLM-as-a-judge sampling pipelines, and trace analysis to hill-climb toward production readiness.

Notable Quotes

"If you want to know what it's doing, look at it at its thinking."

"It's important to test the negatives; checking if the model didn't do something bad is just as critical as checking if it did the task."

Content References