How to Write Your First AI Eval - SallyAnn DeLucia, Arize AI

By ai.engineer

Categories: AI, Tools

Summary

AI evals are non-negotiable for shipping agents—without them, you can't detect regressions, compare models objectively, or enforce quality gates in CI/CD. Unlike unit tests, LM-as-a-judge evals must evaluate multiple valid outputs by examining agent traces, not just single input-output pairs.

Key Takeaways

  1. Vibe checks don't scale: without evals, you can't prove if a prompt change fixed or broke a use case, leaving you blind to regressions over time as problems compound week-to-week.
  2. AI evals fundamentally differ from unit tests because agents have multiple valid outputs and unreadable logic—you must evaluate against traces (logs showing all agent steps) rather than single correct answers.
  3. LM-as-a-judge is the primary eval pattern: use a secondary LLM to grade specific criteria (QA, relevance, summarization) and provide both labels and explanations, giving reasoning behind each grade.
  4. Quality gates in CI require evals—you need objective comparison of prompt and model versions before shipping to detect which changes actually improve performance for your specific use case.
  5. Observability (traces) is a prerequisite for evals: you must capture all intermediate steps an agent takes to understand its behavior and design meaningful evaluation criteria that reflect real complexity.

Related topics

Transcript Excerpt

[music] >> Um, before we get started, I just wanted to introduce myself. I'm Sally Ann, and I'm head of product at Arize. Uh, been at Arize for like 4 years now. Um, one of the first projects that I owned actually was building our AI agent Alex into our platform. So, I've been dealing evals for a little while now, and uh, even building a lot of our evals product. And so, I think the way that I think about evals is that they're our primary signal for our agents. Um, and that includes all of our evals at Arize. So, my goal today is you'll leave with three things. Um, a clear definition of what an eval is and what it isn't. Uh, being able to write your own criteria. So, I'm hopefully going to be able to provide you with like a little bit of a recipe uh, for how you can think about evals and h…

More from ai.engineer