UX audits with parallel agents
You can audit a prototype with several agents at once, each one playing a different persona. The hard part isn't running it, it's deciding who audits.
Auditing your own screen has two flaws: it's slow and it's biased. After forty screens I think everything looks great, because I've piled up context and lost the sense of strangeness. That's confirmation bias doing the QA work. The way out is to not audit alone.
Many heads at once
I spin up the prototype on a local server and fire off agents in parallel, each one with a persona: the classic heuristic evaluator, the user in a hurry, the person who can't see low contrast. Each one scans the screen through its own lens and hands back a report. At the end, a consolidated score that shows where it actually hurts, not where I think it hurts.
The trick is the persona
The part that looks like work, running the audit, that's the AI. The part that matters, deciding who audits, is mine. A shallow persona gives back a shallow report. "Audit the usability" gets me generic. "50-year-old driver looking at the app parked in the lot, in a rush to get out" gets me the real problem, the way it actually shows up for the person using it.
Fresh eyes on purpose
It works because each agent comes in without the context that blinds me. It's the opposite of what happens to me alone: I keep adding context until the screen looks obvious. They start out finding it strange, which is exactly the state the user arrives in. The heuristic became a commodity. What's still mine is knowing which eyes are worth looking through.
// Takeaway AI runs the heuristic in parallel and with no fatigue bias. But the quality of the report is the quality of the personas I wrote before invoking them.