The case
Demo · ILIO Saved to theagent-gateway-evals dataset — it runs from that night on
Its id in every chart: — — fixed once saved.
Name at least one. An open request is answered with “which hotel?” and the turn is spent asking.
One message. There is no follow-up turn, and nothing is staged: no tool is made to fail, no record is planted, the date is today.
One assertion. Two requirements joined by “and” still score a single true or false.
Who to ask when this case starts failing.
Testing needs a prompt and an outcome. Saving needs a name and a squad.
Test runs
Graded by the evaluator the nightly uses — these runs score nothingNo runs yet. A turn takes about a minute and drives the deployed agent for real.