A Hacker News thread last Tuesday had somebody bragging about piping their entire codebase into Claude over MCP. Forty thousand tokens loaded up. And the model still asked, every single message, whether the bug lived in auth or billing.

So window size was not the bottleneck. The model just had no clue which file was on screen at that exact second.

That’s the AI context problem, and a bigger window does not fix it.

A stranger in row 47

Picture a brilliant consultant who lands at your desk knowing your industry cold, but with zero idea you’ve got nineteen tabs open, your CRM is mid-edit, and the typo she’s supposed to fix sits in row 47 of a spreadsheet she cannot even see. By the time you’ve explained row 47 to a stranger, you’ve fixed it yourself, eaten a granola bar, and forgotten what you were doing.

So you compensate. You paste, you re-explain, you build little context-rebuilding incantations at the top of every chat. A five-minute task quietly bloats into ten because the prep work eats most of the savings. We covered the financial side in the AI context tax post. But the cognitive cost is louder. And every fresh chat starts from absolute zero, with your working memory doing the job the model was supposed to be doing for you.

Bigger windows aren’t it

Most people misread this. They keep waiting for GPT-6 or Opus 5 or some unannounced thing with a context window the size of a small moon. So horsepower must be the bottleneck. Wrong. But the model is plenty smart already.

Two prompts, same model. First prompt: “Reply to this email politely declining the meeting.” The model has no clue which email, which meeting, or who sent it, so it spits out the polite mush you’d get from any chatbot anywhere on the planet. Second prompt: an agent sitting inside the browser tab where the email is already open, quietly reading the thread, the prior replies, the calendar invite, and your sent folder for tone. And the reply actually sounds like you wrote it.

Same model. Different context. That’s the whole trick.

Why the browser already won this

Almost every scrap of context an AI needs to be useful is sitting in a browser tab right now, authenticated and rendered and patiently waiting for somebody to ask. Your inbox is signed in. The CRM lives behind a single-sign-on session that took IT two months to provision. Calendar is laid out for the week. But none of that exists in whatever cloud your chatbot is dialing out to, and forklifting it there would be a privacy disaster, because you would essentially be handing every authenticated session you own to a third party with no business sniffing them.

Dassi lives in the side panel of the Chrome you already use. So it just sees whatever tab you happen to be on, signed in as you, without uploading anything to a remote server first. The model under the hood is the same one ChatGPT or Claude would call from their own apps. Only the context shifts. And that single shift flips the output from generic chatbot mush into something that fits your Tuesday afternoon.

Nyne raised real money this spring to give AI agents what they call “human context.” We wrote a longer take. Short version: most teams chasing this are reinventing what the browser already solved, badly, and with venture funding.

Benchmarks lie

Benchmarks grade models on problems some clever PhD wrote down and tidied up first. Real work isn’t like that. Real work is half-finished and scattered across six browser tabs, dependent on a hundred bits of unspoken state you couldn’t articulate cleanly even if a well-funded research lab paid you handsomely just to try. So the same model graded in a vacuum will always look sharper than that exact model dropped into your real Tuesday afternoon, mid-deal-review, with a toddler shrieking in the next room.

And the labs chasing benchmark wins are optimising for a fictional user. But browser agents aim at whoever has to file the damn expense report by Friday.

The quieter half

A model that sees what you’re looking at also knows what to ignore. And the second half matters more.