With Jev
No agent run recorded yet.
No agent run recorded yet.
Without Jev
No agent run recorded yet.
No agent run recorded yet.
webdriver baseline, which drives the page through Playwright, and its webmcp-native treatment, which is handed the tools the page registers. Both run on claude-code with Claude Sonnet 5. They are deliberately not the arena's desktop arm, which has no frame stream to show you.The two Jev arms run read-only against a site that is not ours: each step is judged before it is taken, and one that looks consequential - placing an order, sending a message, deleting something - is refused rather than pressed. An arm that stops that way is labelled Refused (read-only), which is the gate doing its job and not a failure of the page.Jev picks, it never writes. Text to type is chosen from the words of your request. When the request leaves out a value a field needs, a search term, a city, a quantity, both Jev arms ask a small Claude model (Haiku 4.5) for one. A link counts as such a value when the request names or describes a site, so a field asking for one can be filled. It is never asked for identity or account data: a name, an email, an address, a card number stays empty and the arm stops as Partly done. A step that used it is marked generated, and the model's time and cost are shown as their own part of the line, apart from Jev's.The agent arms are capped: three launches per visitor per day, and a daily ceiling across everyone. Past either, you are shown the last comparison recorded for that page and request instead of a new one, labelled as a recording.