Give a model a prompt that says disclaimers can never be removed, say it twice, and it will still find an edge case where it detects the disclaimer and removes it anyway. That failure is invisible in a pass-rate number, and it's the story at the center of "How Evals and Prompts Shape Agent Behavior," a talk by Preetika Bhateja and @DanielJBump of Google on the evals they built for YouTube ads. @aiDotEngineer has it on YouTube. What you get from it is a working order of operations for building an eval system, from vibes to scaled raters. - Fix the tools before you build the eval. A focused set of LLM-friendly tools is the foundation; optimizing those comes before large agent evals. - Vibing early is a feature. Intuition-based looking at outputs is non-scalable, but it surfaces failure patterns fast and lets you make radical architecture changes without an eval holding you back. - Jumping to scaled raters too early hurts. You end up calibrating the eval and changing the model at the same time, and the numbers swing. - Start small and test the negatives. No massive golden set on day one. Checking that the model didn't do something bad matters as much as checking it did the task. - The rubric is the hard part. Writing the eval is small; humans arguing over what the rubric should be is the big work. Human-human agreement inside your team has to be strong first. - Ask raters for explanations, not just pass or fail. Especially for multi-output cases where an ad is brand safe but inaccurate, the reasoning tells you where the agent is actually missing. - For LLM judges, monitor the disagreement rate. Sample how a human expert rates against how the model rates, and watch the trend. - Look at the traces. The disclaimer removal only showed up in the agent's thinking, never in the categorical pass rate. - Chase patterns, not single runs. These systems are non-deterministic, so patching a prompt off one failing trace is a trap. - Agents don't generalize. Same as traditional ML: keep an edge-case set and a test set, use it sparingly, refresh it with production data. - Decide your launch gate up front. Know which regressions are acceptable trade-offs and which are critical failures before you're staring at the ablation. I'm working through the published talks from AI Engineer World's Fair sharing summaries and takeaways. Follow for more!