AI testing

Agents cover the breadth.
People make the calls.

Controlled agents and automation run inside a scope a person wrote and you signed. A senior reviewer confirms every finding by hand.

Written scopeStop conditionsStaging by defaultConfirmed by hand

The agents

What the agents do.

Agents walk the agreed scope the way a user does — click, type, submit. The work is dull on purpose. Breadth. Repetition. Patience.

Breadth

An agent walks every page and flow the scope names, like a user with endless time.

It goes where the scope says, and nowhere else.

click · type · submit

Repetition

Agents try the boring permutations people skip — each field, each state, each order of steps.

No run gets less attention than the last.

each field · each state · each order

Patience

Agents record everything they see — screens, requests, timings — as evidence for the review.

Nothing rests on memory.

screens · requests · timings

The automation

What the automation does.

Some checks should never depend on judgment or mood. Those run scripted, the same way every time.

Runs the same, every time

  • Security headers, present and configured
  • TLS across the tested scope
  • Redirects that land where they claim
  • Broken flows — links, forms, dead ends
  • Web vitals on the journeys the scope names

When a check runs the same way every time, a change reads as a change. The value is sameness.

The people

Where people decide.

A person writes the scope. Agents run inside it, under stop conditions. People decide what reaches the card.

  1. Run

    Agent

    Walks the agreed scope like a user.

    stop conditions apply

  2. Record

    Evidence

    Screens, requests and timings, as seen.

    handled as the scope says

  3. Judge

    Human review

    A senior reviewer confirms each finding by hand.

    false alarms die here

  4. Sign

    Report card

    Only confirmed findings reach it.

    found by agent · confirmed by hand

The scope
A person writes it and you sign it. Agents never pick their own targets.
Stop conditions
The scope names where an agent stops. Anything outside the scope stays untouched.
The review
A senior reviewer confirms each agent finding by hand before it reaches the card. False alarms die here.
The signature
A person signs the report card. A tool's confidence score never does.

The rails

The rails around every run.

The rails are written before anything runs. You can check every one.

Before an agent runsagreed in writing
  • Scope + consentwritten and signed before anything runs
  • Environmentstaging by default · read-only on production, when the scope says so
  • Accessleast access — only what the scope grants
  • Credentialsrotated when the test ends
  • Evidencehandled and shared as the written scope says

The written scope carries every rail.

Limitations

What agents miss.

An agent can cover a scope and still miss the point. Meaning. Taste. Context.

Meaning
A flow can pass every check and still confuse the person holding the phone.
Taste
A page can work and still feel cheap. No agent notices.
Context
What your product promises, who it serves, what a miss would cost — an agent carries none of it.

That is why people grade, and why a person signs the card. The signature is the judgment.

signed by a person

Next

Get your report card.

See how agent findings arrive on the card, or send a brief. A URL and two sentences is enough.

Send a URL, never credentials or secrets.

Report card in 72 hoursNDA + written consentFound by agent, confirmed by hand