The HN front page today has a piece called “Beware AI Productivity Theater” sitting at #11, and the comments are doing what HN comments do, which is to say agreeing furiously while rephrasing the same complaint forty different ways. The gist: AI demos look incredible on stage and then evaporate the moment you point the same tool at the work that pays your salary.

So the diagnosis is right. But most of the thread missed the cause.

The demo always works in a fresh tab

Watch any agent demo from the last six months. The agent books a flight on a sample site. It scrapes a public dataset. It summarizes a Wikipedia article. Beautiful. So then you open Salesforce, or your company’s Jira, or your accountant’s filthy PDF portal, and the agent stares at the login screen like a confused tourist.

And this is not a model intelligence problem. The model is fine. GPT-5.2 and Claude Opus 4 can both reason about a Salesforce opportunity record if you paste the HTML at them. But the agent has no way to be inside the place where your actual work is happening.

What “authenticated context” really means

Your job lives inside maybe ten browser tabs. Gmail. Notion. Linear. Stripe. HubSpot. A spreadsheet with the numbers your boss cares about. A second spreadsheet with the same numbers but worse formatting because someone in accounting hates you. Each of those tabs is held together by a stack of session cookies, single-sign-on tokens, sometimes a hardware key check, occasionally a CAPTCHA challenge that fires if anything sniffs around in a way the site doesn’t like. So the state that makes those tabs yours is not portable, and you cannot zip it up and ship it to a cloud agent running on a VPS in Oregon.

So when a cloud-hosted agent says it can “use your Gmail,” what it really means is one of three things: you handed it an OAuth token and now it has lifetime read-write access to everything in your inbox (good luck revoking that cleanly later), or it pretends to support your service via a brittle API integration that breaks the next time Google ships a UI change, or it spins up an entirely new browser somewhere far away and asks you to log in again, this time inside a sandbox you neither own nor can audit.

And none of these three options is the same thing as the browser tab you already have open.

A small example from last week

I asked an agent to pull the last twelve months of expense data from a SaaS billing dashboard. The dashboard had no API. The cloud-hosted agent could not see the dashboard. So it offered to “guide me through it,” which is what AI says when it’s giving up. Then I opened Dassi in the side panel of the tab I was already logged into. It read the table, paginated through twelve months, dumped the result into a CSV, and stopped. Eight minutes. No reauth, no token handoff, no fresh sandbox.

So that’s not magic. The agent ran in the same Chrome session I was already using. And the session was the entire trick.

Where the HN thread gets warm but not hot

A lot of the productivity-theater critique frames the issue as “models hallucinate” or “agents can’t plan multi-step tasks.” Both true. Both fixable over time, probably, with better tool use and longer context windows and the usual incremental model improvements. But there’s a deeper structural problem that no amount of model improvement touches: if the agent lives in a server farm and your authenticated work lives in your laptop, those two are separated by an air gap that OAuth tokens and remote-browser sandboxes do not bridge cheaply or safely.

And I wrote about the same idea from a different angle in the context tax piece and again in the cloud-vs-local browser post.

The HN commenters who get it usually phrase it the same way. They say something like “I’ll believe AI productivity when my agent can do my actual job.” Yes. That’s the right test. But the only way it passes is if the agent sits in the same Chrome window where your job lives, looking at the same DOM you’re looking at, using the same cookies your browser is already managing.

A two-line aside

So most cloud agents are not bad products. They are good products pointed at the wrong layer of the stack. The layer where work happens is your authenticated tab.

Theater vs. output

If you can’t show me an agent doing a real task in a real tab against a real logged-in service, you are showing me theater. It might be pretty theater. The lighting might be great. But it is not work.

Dassi runs as a Chrome side panel agent inside your existing browser session — install link. No OAuth dance, no remote browser, no token vault to maintain. And we did not pick this architecture because we’re smarter than the cloud agent folks. We picked it because it’s the boring one, the one where the agent sits in the place your work already happens.

The HN post will fall off the front page tomorrow. The structural problem it points at will still be sitting there.