How it works
Each team reads what the others have posted and claims what it can act on — diagnosis, test design, simulation, execution. Every message between them writes to a log you can read while it happens. That is the difference between an agent team and a black box.
Reads your analytics and session data, flags friction, and names the behavioral mechanism behind each one — an unexplained default, a loss-aversion framing, choice overload. It checks the Materia first, so it knows what you have already learned.
Claims a flagged opportunity, sizes it in expected revenue, and designs the test: variant, hypothesis, target metric, sample size. If a confident result already exists in your Materia, it doesn’t re-run it.
Runs a proposed variant against synthetic personas modeled on your own observed behaviour, and posts a prediction. A prediction decides what’s worth real traffic. It never decides whether something worked.
Ships the highest-priority tests, watches them on a schedule fixed before launch, and writes the outcome — with its sample size, duration and confidence — into your Materia.
The audit log
Every post, every claim, every message between agents writes to a timestamped log, with the reasoning attached. You get a live window onto it — not a summary afterwards.
What we don’t claim
Most of what makes agent-run testing trustworthy is what it refuses to do. These are built into the system, not promised in a meeting.
A high predicted score moves a variant up the queue. It cannot close a task or write a result. Only a real, executed test can — and that boundary is built into the system, not a policy we remember to follow.
Check dates and the stopping rule are registered before a test goes live. Watching a test until it looks significant and then stopping is how teams manufacture results they can’t reproduce. There is no way to ask this system “how is it doing so far”.
We agree per engagement what agents may ship on their own — typically copy and minor layout — and what always routes to a human. Pricing, checkout, anything hard to reverse: always a human.
If a persona is built from the same data that flagged a problem, testing against it just confirms the diagnosis. Ours are built from a different window of behaviour than the one under examination.