Stop shipping AI agents on vibes: evaluate what you can't predict

The problem. Most AI agent teams ship with "looks good to me" as their test suite. Laurie Voss demonstrates this pattern: you try a few inputs, they look right, you ship, and then a customer finds a broken edge case. The underlying issue is that LLM outputs are nondeterministic, so unit tests with expected strings don't work. The video proposes a systematic alternative: traces and evals.

The solution stack. Voss presents three layers of evaluation: code evals for deterministic checks (ticker mention, JSON parse), LLM-as-judge evals for meaning (accuracy, faithfulness, actionability), and human evals for gold-standard labeling. The key insight is that no single method catches everything; you need all three in a Swiss-cheese model. Phoenix, Arize's open-source observability tool, automates much of the plumbing.

The trap of the wrong eval. The video's most useful moment shows that the built-in correctness eval fails completely because it judges against the model's knowledge cutoff. Switching to faithfulness – which checks if the report is grounded in the agent's actual research – produces far more useful results. The lesson: choosing the right question beats tuning the prompt.

Testing your tests. A common failure is building a single "god evaluator" that checks everything. Voss recommends splitting into many small evals, each with a clear dimension. He then shows how to measure judge accuracy against a human golden dataset, using precision and recall. Even with small samples (16 reports), you can detect disagreement and iterate on the rubric.

The eval-iterate cycle. Once evals are trustworthy, you can run controlled experiments. Voss demonstrates improving prompts based on eval findings, and then comparing results before and after. He upgrades the model from Claude Haiku 4.5 to 5.5 and finds that actionability improved but faithfulness remained poor – and the model became overly conservative, mostly saying "hold". The cycle of measure, change, remeasure is where the value lives.

The organizational implication. Voss argues that anyone who knows what good looks like – product managers, support, sales – can write evals by describing criteria. This shifts evaluation from an engineering-only task to a cross-functional capability. The final piece is putting regression evals into CI, gating builds on guardrails (ticker mention) and north stars (actionability average). This transforms AI agent development from vibe-driven to evidence-driven.