Golden Sets That Do Not Rot: Building a Test Suite for an AI Agent

An AI agent is only as trustworthy as the test it has to pass, and that test is its golden set — a curated collection of questions with known-correct answers. Build it carelessly and it measures nothing while looking rigorous. Build it well and it becomes a durable definition of 'good enough' that catches regressions for years. The difference is entirely in the curation and the upkeep.

What a golden set is for

A golden set answers one question: is this agent actually correct on the things it's supposed to handle. It's not a vibe check or a demo; it's a set of real questions from the agent's domain, each paired with the answer a domain expert agrees is right. When the agent is gated on it, the golden set is the wall between 'seems fine' and 'shipped' — which means the wall is only as good as the set.

The two ways golden sets fail

• Too easy. A set of softball questions the agent was always going to get right measures nothing and provides false confidence. If everything passes on the first try, the set isn't testing — it's flattering.
• Too stale. A set written once and never touched slowly drifts from what the agent actually faces, until it certifies competence on a job the agent no longer does.

Both failures look like success — green checks all around — which is what makes them dangerous.

Curation is the craft

A good golden set is built from real cases, not invented ones, and it deliberately includes the hard ones — the edge cases, the ambiguous inputs, the questions where a plausible-sounding wrong answer is easy to give. Each answer is verified by someone who actually knows the domain, because a golden answer that's subtly wrong is worse than no test at all. The work is unglamorous and it's the whole game: the set's value equals the care that went into it.

Keeping it from rotting

The defense against staleness is treating the golden set as a living, versioned artifact. When the agent fails in production in a new way, that failure becomes a new golden case, so the set grows toward the real failure modes over time. When the domain shifts, the set shifts with it. Versioning matters too: you should be able to see how the set changed and re-run an old version, the same way you'd manage test code. A golden set that's maintained gets sharper; one that's frozen gets blunter.

Golden sets across the lifecycle

The same golden set does double duty. It's the publish gate on day one, and the tripwire for continuous evaluation after launch — re-run on a schedule to catch the day the agent quietly degrades. One well-curated set protects the agent at ship time and throughout its life, which is a strong argument for investing in it properly rather than assembling it in an afternoon before launch.

Build the exam that actually tests. See how a well-curated, versioned golden set gates an agent at publish and catches drift after launch — one investment that protects an agent for life. Book a walkthrough.