Reliability
Eval
A repeatable test of whether your AI system is actually getting things right — the difference between improving it and guessing.
Traditional tests are deterministic: the same input gives the same output. AI systems are not, which is why teams so often ship a feature and then have no idea whether their next prompt change helped or hurt.
Evals fix that. For agents the highest-signal test is usually not "was the prose good" but "did it call the right tool with the right arguments" — an assertion that is cheap, fast and runs like an ordinary test.
A workable progression: collect real production examples, categorise how they fail, turn the failures into test cases, and run them on every change.
Without this, teams eventually stop touching the system out of fear. That is a worse outcome than any single bug.