PlateTaleAI agents can browse the web. Why can't they reliably use it?
Browsing and using are different problems. Reading a page is close to solved. Acting on a page — reliably, in sequence, with real consequences — is not.
Reading is one step. Using is many.
A ten-step journey with 95% reliability per step completes about 60% of the time. The individual steps look fine; the journey does not. Reliability compounds downwards, which is why demos succeed and production runs fail.
The web gives no feedback contract
An API returns a status code. A website returns a new picture of itself and leaves you to work out what changed. Agents have to infer success from re-reading the page, and inference is where mistakes enter — repeating a submitted action, or declaring victory on a page that quietly rejected the input.
The environment is adversarial
Bot defences, rate limits, region redirects, A/B tests and consent walls all exist to shape human traffic. To an agent they are non-deterministic obstacles that appear on some runs and not others.
What closes the gap
Not better models alone. Sites need to expose state clearly and offer explicit paths for automation — and someone has to keep measuring whether the journeys actually complete. That measurement is the part almost nobody is doing yet.