Home / Platform

A voice agent can pass every integration test and still lose the caller. It answers the wrong question confidently, talks over an interruption, forgets what it was told two turns ago, or handles an emergency like a billing query. None of that shows up in a green build. Prologue is the QA layer for that gap: it evaluates long-form, unstructured conversations, runs them against the real product on web and mobile, and turns the failures into tests that block the next regression.

Three parts, one pipeline

• Evals Latency, tool-call accuracy and time to first response, scored in the same run as personality, patience and how the agent handles an emergency.
• CI/CD for web & mobile Drive the real product: iOS and Android UIs alongside web, so the agent is tested through the surface your customers actually reach it on.
• Personas & custom tests Every persona and edge case gets a suite of its own, wired into your pipeline so it runs on each deploy, next to prompt and code review.

Why voice agents need their own QA

Text-based evaluation misses most of what goes wrong on a call. Transcription errors change the meaning of a turn before the model ever sees it. Barge-in and turn-taking decide whether a caller feels heard or interrupted. Latency that reads fine in a log feels like dead air on a phone line. And the same reply can be correct and still land badly with an anxious caller.

Prologue scores those dimensions together, so a change that improves tool-call accuracy but makes the agent curt shows up as the tradeoff it is, before it ships.

Where it runs

Prologue drives the product end to end rather than calling an endpoint: the web app, the iOS app and the Android app, with the same suites across all three. Runs happen on demand while you iterate on a prompt, and automatically on every deploy once they are wired into CI.