# Trust Lab

Break it in the lab. Not in production. Turn real conversations into scenarios the agent must always handle correctly, run them on every change, and catch regressions before a single user does. Early access. Works on top of the Product Agent.

## Lock in the behavior your users depend on

1. **Import real conversations** — pull the conversations that matter from production or the Playground. Evals and mock results are auto-generated from what actually happened.
2. **Define what correct looks like** — refine the evals on each turn: what the agent must say, which actions it must call, where it must navigate.
3. **Run on every change** — group scenarios into suites, schedule them daily, and watch the dashboard. A falling pass rate is your regression alarm.

Playground is where you experiment. Conversations show what users did. Trust Lab is where correct behavior gets locked in.

## Every failure, explained

A failed run shows each turn with the agent's actual response next to the eval that judged it — what was expected, and what the agent said instead.

## Features — define correct, detect drift

- **Start from real conversations** — import from production or the Playground, searchable by user, message, or action. Evals and mock results are auto-generated.
- **Author multi-turn scenarios** — build test conversations from scratch: script each turn and attach the evals that define a correct response.
- **Grade what the agent says** — LLM-as-Judge for nuanced criteria in plain English, plus Contains, Exact Match, and Regex for fast deterministic checks.
- **Grade what the agent does** — tool-use evals verify the right action called with the right parameters, the right route navigated, the right knowledge retrieved.
- **Test as any user** — set the simulated user's plan, role, and locale, so plan gating, role-based behavior, and localized responses get tested.
- **Organize into suites** — group scenarios by journey or feature area; suites are the unit you run, schedule, and track.
- **Run on a schedule** — hourly, daily, weekly, or custom cron, in the timezone you pick. Regression detection runs whether anyone remembers to or not.
- **Drill into every run** — every execution logged with outcome, pass rate, and trigger. Open a failed run to see the exact turn, the eval that failed, and what the agent actually said.
- **Never ship the same bug twice** — when a real conversation goes wrong, import it. It becomes a regression test that ensures the same mistake can never happen again.

## What teams lock in with Trust Lab

- **Ship changes without fear** — change a prompt, an action, or the knowledge base, then rerun every suite and see what it did to quality before users feel it.
- **Swap models, keep quality** — prove a faster or cheaper model holds the line on your own scenarios, not a public benchmark.
- **Catch drift overnight** — knowledge updates, API changes, model updates: a falling pass rate is your signal that something the agent used to get right has broken.

Trust Lab runs the tests. You ship knowing what still works — and what just broke.

Book a demo: https://cal.com/team/foldspace/demo