In this post, we walk through the full workflow with a financial market-intelligence agent. We capture failures from production traces, build a versioned dataset, run an evaluation, fix the agent, and confirm the improvement against the same locked inputs.

Agents are non-deterministic by design. The same input can produce different outputs across runs, which makes a single evaluation result nearly meaningless. You can’t tell if a score moved because the agent changed or because the model sampled differently. Consistent measurement across stable inputs is the only way to know whether a change actually helped.

But stable inputs alone aren’t enough. A large language model (LLM) judge can tell you whether a response sounds helpful. It cannot tell you whether the stock price is accurate, whether the broker workflow ran in the right order, or whether personally identifiable information (PII) leaked between sessions. For those checks you need ground truth: the expected response, the required tool sequence, and the assertions that must hold regardless of how the response is phrased. Ground truth is what turns a subjective score into a verifiable measurement. Without it, you’re measuring the appearance of correctness, not correctness itself.

Versioned datasets give you both. They hold the inputs still so scores are comparable across runs, and they carry the ground truth that makes those scores mean something. This matters most in the two places where agent evaluation actually happens.