Engineering
7 min read
Evaluation first: how we test agents before we trust them
Written by
Ashgrid Labs
Golden sets, adversarial cases and human review loops — the testing stack behind every Ashgrid deployment.
Write the tests first
We build the evaluation suite before the agent: a golden set of real cases, a set of adversarial ones, and a clear definition of a pass.
Golden sets
Fifty to two hundred real examples, labelled by the people who do the work today. This is the bar the agent has to clear.
Adversarial cases
Ambiguous inputs, missing data and edge cases your team has seen once a year. If it fails these quietly, it isn’t ready.