Home / Platform / Evals

What gets scored

An eval run takes a set of conversations, replays them against your agent, and grades the result on every dimension you care about at once. The technical half is the part most teams already log:

• Latency and time to first response. What reads as a fine p95 in a dashboard can still feel like dead air on a phone line.
• Tool-call accuracy. Whether the agent called the right function, with the right arguments, at the right point in the conversation.
• Task completion. Whether the caller got what they called for, not whether the turn was well formed.

The half that usually goes unmeasured is the one callers actually notice:

• Patience and tone. Explaining something the fifth time as willingly as the first.
• Recall across turns. Whether the agent still knows what it was told earlier in the call.
• Escalation judgement. Recognising an emergency and treating it like one rather than like a billing query.

Graded on long-form conversation, not single turns

Most evaluation tooling scores one prompt and one completion. Voice calls are neither. A failure at turn twelve is usually caused by something the agent did at turn three, and a rubric that never sees the whole call cannot attribute it. Prologue grades the conversation end to end, so the score points at the turn that caused the problem.

Judges you can check

An LLM judge that nobody has validated is a number, not a measurement. Rubrics are built from your own transcripts, checked against human labels on a held-out set, and reported with their agreement rate, so you know how much to trust a score before you gate a release on it.

Where evals fit

Run them by hand while iterating on a prompt, then wire the same suites into your pipeline with CI/CD testing for web and mobile so they run on every deploy. Failure modes you find here become personas and custom tests that keep them from coming back.