Skip to main content
Grading live traffic tells you how agents did on the work that happened to arrive. Golden Scenarios answer the other question: does this agent still handle the case we already know it must? They are versioned test cases with a known outcome, run against an agent on demand or on a schedule, and their verdicts land in the same score ledger as organic grading. This page covers what a scenario holds, how versions move, and how to read a result. Find them under Quality Lab → Golden Scenarios.

What a scenario holds

A scenario has an identity — name, domain, description, tags — and its content lives in versions, so you can change a case without losing what it used to say. Each version carries:

Versions move through review

draft → validating → awaiting review → published → archived The first three are reversible — a version can go back for more work. Published and archived are one-way, so a case other people’s results depend on cannot be quietly rewritten underneath them.

Suites group scenarios against an agent

A suite is a named group scoped to the agent it tests. Membership carries two things beyond the scenario itself: whether it is enabled, and its weight — so a suite can hold a case temporarily switched off, and can count the cases that matter more for more.

Reading a result

Each run records the scenario, the version, per-phase results, a score, and a failure kind. Two outcomes look similar in a list and mean different things:
Ungraded — the case ran to completion but was never judged. The work happened; the verdict did not. Most often the judge could not be dispatched.Failed — the case never completed at all: it failed, was blocked, or was cancelled. It was never gradeable, so there is no score to be missing.
Keeping them apart matters when you read a suite’s summary: ungraded cases are a grading problem to chase, failed cases are an execution problem. Golden verdicts are written into the same score ledger as organic evaluations, so an agent’s scorecard shows both together and you can see whether a regression suite and live traffic agree. A scenario can also be brought in from a bundle rather than authored in place, which is how a set of cases moves between workspaces.

Next steps

Evaluations

Grading the traffic that actually arrived.

Model Evaluation

Comparing a candidate against production on the same input.

Continuous improvement

Turning verdicts into a proposed change.

Runs & evidence

A scenario run is a run — open it and read it.