Three separate threads landed on Hacker News this week asking the same damn question. “Why aren’t AI productivity gains higher?” One was an Ask HN about whether companies should fire developers or build better products. Another picked apart the AI productivity paradox. And a third made the case that AI actually creates more work. Meanwhile, Yamak and Skyvern 2.0 both shipped new browser agent products within days of each other, and both of them want you to configure credentials before they will do a single useful thing.

I keep thinking these are the same conversation, just split across tabs.

Thirty-four minutes

Last month I tried setting up a browser automation agent. Not naming names because this applies to basically all of them. The marketing page said “automate any browser task in seconds,” so I installed it, created an account, generated an API key, pasted that key into a config field, granted OAuth access to my Google account, approved three separate permission scopes, set up a proxy configuration for authenticated sessions, and then got stuck on a CORS error that required yet another browser extension to fix. Thirty-four minutes before it completed one action.

And I do this professionally. A product manager would have closed the tab at step three, gone back to copy-pasting between windows, and quietly filed this under “AI overpromises.” Because the surveys measuring productivity are not capturing model capability. They are capturing the entire experience, including the miserable setup gauntlet, and that is where most tools lose people before anything intelligent even happens.

Why cloud browsers make this worse

Skyvern’s new release looks impressive for structured data extraction, I will give them that. But their cloud browser boots from a blank session, which means you authenticate into every service from scratch, babysit session tokens that expire on their own schedule, and re-authenticate when something breaks at 2am on a Tuesday. Okta says the average knowledge worker touches 89 SaaS apps. Multiply the authentication overhead across all of those and you have built yourself a part-time job maintaining plumbing that was supposed to save you time.

This is the core problem with cloud browser agents. They spin up a Chrome instance that knows nothing about you. No cookies, no saved passwords, no sessions. So every single task begins with authentication, and the minutes lost there quietly eat whatever the AI would have saved.

Already solved, just ignored

Your browser already has all of this figured out, which is the part that keeps getting lost in the HN discourse. Gmail is logged in. Slack is logged in. Your CRM, your project tracker, your weird internal admin panel that only works in Chrome. You did the OAuth dance months ago.

Dassi sits in your browser’s side panel and inherits that entire authenticated state without asking you to configure anything, because there is nothing to configure when the sessions are already right there in the browser where you are working. The gap between wanting help and getting help drops from half an hour to maybe four seconds. That difference is not a feature comparison bullet point. It is the difference between a tool people actually use and one they bookmark with good intentions and never open again.

GPT-5.2 does not care about your OAuth scopes

The AI productivity paradox was never really about models. GPT-5.2 and Claude Opus 4 can reason through complex multi-step workflows, parse messy web UIs, and draft emails that sound like a human wrote them. The capability is genuinely absurd at this point. But none of that matters if the user bounces during setup, and from what I can tell, most users bounce during setup.

Browser agent companies keep shipping incredible AI on top of onboarding flows that filter out 90% of potential users before the model ever fires a single inference. And then we all sit around reading blog posts about why productivity gains have not materialized.

The gains exist. They are sitting behind a credentials page that most people never finish filling out.