Light split into clean bands through a prism in the dark

Open books

How we test
every agent

Most companies ask you to trust the AI. We would rather show you the tests. Every agent rehearses before it goes live, red-teams try to break it, and the whole thing runs green before anything reaches a customer.

The standard

An agent rehearses before it ever meets a customer.

Every agent has its own rehearsal suite: a scripted set of real situations it has to handle correctly, from the easy ones to the ones designed to trip it up. The agent cannot go live until it passes the whole suite.

That is the shared bar, not a per-project favor. And it does not stop at our desk: every client gets a fresh live re-run of the suite against their own setup, their own facts and their own rules, before go-live. The rehearsal that matters is the one on your account, not a generic demo.

We write these suites by hand, the same way we write the agents: small, readable, and run on every change.

By the numbers

Real counts, computed and dated.

We would rather show a smaller true number than a rounder invented one. These are counted straight from the codebase. When they change, the date changes with them.

17
agent rehearsal suites, one per agent
60
automated test suites guarding the code
3,749
individual checks, all passing green

As of July 19, 2026. Counted by listing the rehearsal suites and the automated test suites, then running the fast deterministic suites and summing every check each one reports as passed.

Red-teams

We try to make the agents misbehave.

Passing the friendly cases is the easy part. The suites that matter are the ones that attack the refusals, the moments where a careless agent would do real harm. An agent has to hold the line on all of these to ship.

  • Invented prices

    It must refuse to name a price or a fact you never gave it. Made-up numbers are a failure, not a rounding error.

  • Missed opt-outs

    STOP, unsubscribe, "no thanks": the agent has to catch the word and stop, every time, however it is phrased.

  • Cross-client leakage

    One business's agent must never surface another business's data. The isolation is tested, not assumed.

  • Prompt injection

    Instructions hidden inside an email, a form or a web page get ignored. The agent follows only the rules you set.

When a test fails

The change does not ship. Full stop.

Our systems fail closed. If a check cannot pass, we stop and fix it rather than push it live and hope. A red suite is a blocked release, not a warning we wave through.

That is the whole point of testing in the open: the tests are not decoration, they are the gate. Nothing customer-visible reaches your customers on a maybe.

No agent meets a customer before it passes its rehearsal.

Want to see the rest of how we handle your data and your customers' trust? The security page lays it out in plain words.