AI agent testing
The tests that needed a human now run every night.
Not a rule engine with better marketing. Agents operate your site the way a tester does — and hand you the transcript so you can audit the test instead of repeating it.
The fleet
Eight agents, one protocol
-
Keyboard & focus
Tab order, visible focus, traps, skip links, focus after dialogs and route changes. Output: the keystroke trail with a verdict per step.
-
Screen reader
The page read with NVDA, JAWS and VoiceOver semantics. Output: the captured speech stream against what should have been announced.
-
Alt-text relevance
A vision model asks whether the text describes this image, in this context — decorative, informative or functional. Output: a per-image judgement and a proposed alt.
-
Zoom, reflow & contrast
400% zoom, 320 px reflow, text spacing, contrast in real rendered state including hover and focus styles.
-
Forms & errors
Labels, instructions, error identification, error suggestion, required-field handling and whether errors reach assistive technology.
-
Media & captions
Captions present and synchronised, audio description, autoplay, controls reachable by keyboard.
-
Pointer & voice control
Target size, dragging alternatives, pointer cancellation, and whether visible labels match accessible names for voice control.
-
Content sense
Link purpose out of context, heading outline, sensory-only instructions, language of parts.
Evidence
Every finding shows its work
A score you cannot interrogate is a liability. Open any finding and you see the run that produced it: the steps taken, what was expected, what happened, and the WCAG criterion it maps to. Auditors and developers look at the same artefact.
| Agent | Evidence kept | What a human does with it |
|---|---|---|
| Keyboard & focus | Keystroke trail, focused element per step, screenshot at failure | Reproduce in 20 seconds, or confirm and hand to a developer |
| Screen reader | Speech stream vs. expected announcement | Judge whether the wording is acceptable in your domain |
| Alt-text | Image, current alt, judgement, proposed alt | Approve, edit or reject the proposed text |
| Forms & errors | Field-by-field check log, submitted state, announcements | Decide whether the pattern needs a design change |
Honest limits
Where the agents stop
Some judgements are not ours to make. These stay with people, and the product says so in the interface rather than quietly guessing.
We publish the agreement rate between each agent and our human auditors, and it is visible in your dashboard. When an agent is wrong, you mark it false positive and the run is retrained against it.
- Plain-language quality and editorial intent
- Sign-language interpretation quality
- Legal sign-off on the accessibility statement
- Flows requiring real payment or real identity documents
- Anything below the confidence floor — it goes to the “needs a human” queue
Point an agent at your hardest page
Checkout, booking flow, application form — pick the one your team dreads testing.