A Scorer is how you decide whether an agent’s output is good. AGNT5 ships built-in Scorers for the common cases — LLM-as-judge for open-ended outputs, exact-match and structured comparison for deterministic outputs — and supports custom Scorers for anything domain-specific.

You’ll find Scorers under Improve → Scorers in Studio.

Scorers are what turn an Experiment from a bag of outputs into a comparable result. Pick the wrong Scorer and you’re measuring the wrong thing; pick the right one and the improvement loop becomes self-directing. Scorer outputs are stored alongside the Runs they scored, so you can always drill from an aggregate score into the individual cases that produced it.

A deeper guide is in progress. For the full improvement loop, see the Improve overview.