The CUA Reliability Gap

Computer-use agents can pass the demo and fail the job, and the distance between those two is wider than it looks from the outside. Ten clicks into driving some legacy desktop app, one looks close to magic. Then it misreads a modal, files a form against the wrong record, and nothing flags it until a human notices something downstream a day later.

The problem isn’t intelligence

It’s that the agent has no reliable sense of whether the thing it just did actually worked. It reaches for the most plausible next action, not the verified-correct one. Give it a smarter model and it reaches for a more plausible wrong action. The gap doesn’t close.

What closes it is a verifier

Something that sits beside the agent, watches what it did, and checks the resulting state against what was supposed to happen. In practice that means diffing the underlying data store before and after, asserting the invariants that have to hold if the task really succeeded, and rolling back or escalating to a human the moment one of them doesn’t. Deterministic software wrapped around a stochastic actor. It’s dull to build, and it’s the part that makes any of this shippable.

Why anyone pays for it

Nobody putting agents into a real workflow wants “usually right.” They want a number they can hand to their own customers and stand behind. That number is what the verification layer produces, and it happens to sit at the point of maximum leverage in the stack, which is a good place to be spending your time.