It’s quite clearly not the same
It’s literally not the same. I’m not really sure what else to tell you.
They are not interacting with Amazon at all. They are sending control commands to your computer, where you've opened a session logged into Amazon as you. Amazon is claiming they can't run commands on your computer.
The whole thing becomes entirely absurd if you consider that a cloud model isn't even fundamental to the setup; you could just as well slot in a local model or codex or anything else and the scenario would be exactly the same. Perplexity is just acting as a model provider here.
When you say I am "manually controlling the site", that actually means I run a program on my computer that makes requests to the site, chooses how to display the content provided by the site, and then gives me control mechanisms (e.g. clicking, scrolling) to interact with that content.
An "agentic workflow" does the same thing. It's a computer program that I run that makes requests to the site, chooses how to display the content provided, and then gives me control mechanisms (e.g. typing text) to interact with the content.
I'm not trying to be pedantic here. I really think that the two aren't that different.
It’s not using Codex to steer a headless browser
"When a Comet user directs the Assistant to locate an item on Amazon.com, the Assistant takes screenshotsof the browser view, sends those screenshotsfrom the user’s computer to Perplexity’s servers, and receives instructions from Perplexity’s servers on how to navigateAmazon.com. In other words, the Assistant cannot operate wholly independently; it relies on direction from the user and instructions from Perplexity’s servers."
As a consumer, I don't think it's very prudent to trust a company with that kind of access to your data, but it doesn't seem materially different from the kind of trust you have to give to Chrome.
It's not longer my agent. It's a joint agent of my self and whoever else is giving it instructions. (This is Amazon's argument.)
Also, Amazon told Perplexity to fuck off. If your assistant is told to stop doing something in a cease and desist, and then keeps doing it, the fact that you asked them to do it isn't a get-out-of-jail-free card.
My guess is we'll find a balance where local models can act as user agents while corporate ones have to meet certain requirements to gain that safe harbour.