An Anthropic researcher gave everyone a peek at self-improving AI this week, and the discourse did the thing it always does: people started sketching the curve upward, models getting better at their own training loop, each generation shortening the runway for the next. I believe the trajectory. What I don’t believe is the assumption quietly bolted onto it, which is that a smarter model automatically becomes a more useful one for the work already sitting in the tabs you have open.

Capability and context are separate problems. Only one of them is on that curve.

The part the curve doesn’t touch

Say the recursive loop works exactly as advertised. Next year’s model reasons circles around Opus 5. It plans better, recovers from its own mistakes better, writes code that doesn’t need three rounds of review.

It still opens in a tab where you are nobody.

Your Gmail sits behind a session cookie. Your CRM sits behind SSO with a hardware key. The admin panel for the internal tool someone on your team built in 2023 sits behind a VPN and a login your ops lead provisioned by hand, and there is no documentation for it, and the person who wrote it left. None of that softens because the model got smarter. It’s a genius temp on their first morning with no badge, standing in the lobby, thinking beautifully about a building they can’t enter.

Intelligence doesn’t authenticate

Last month someone told me their agent “couldn’t handle” pulling last quarter’s closed-won deals out of Salesforce. I asked what the failure looked like. It wasn’t reasoning. The agent had been handed an API token scoped to something adjacent to what was actually needed, so it confidently returned the wrong rows and explained its work in three tidy paragraphs.

A better model does that same task wrong with more eloquence.

Three things that don’t improve with model quality

  • Session cookies. A cloud agent doesn’t have yours, and no amount of scale gives it one.
  • Bot detection. Datacenter IPs get flagged by Cloudflare regardless of what’s running behind them.
  • Permission scope. OAuth grants what the integration asked for years ago, not what you need today.

Your browser already solved this

Flip the order. Instead of giving a remote model credentials so it can reach your accounts, run the model where the credentials already are.

That’s the whole argument for a browser agent. Dassi lives in your Chrome side panel and works inside the session you already have going, so the Salesforce tab it reads is the one you logged into this morning with your key, your permissions, your view of the pipeline. There’s no OAuth flow, no service account, no headless browser somewhere in us-east-1 failing a captcha. You can install it from the Chrome Web Store and point it at whatever’s on screen. I’ve written before about how smarter models don’t fix a context problem, and self-improvement discourse is the purest version of that mistake I’ve seen yet, because it takes the one axis we know how to improve and treats it as though it were the only axis that exists. The badge, not the brain, is the damn bottleneck.

Why this makes the capability jump matter more, not less

Here is the part I actually find interesting. If the model runs where you’re already authenticated, every capability release lands directly on your real work instead of on a demo.

Dassi is BYOK, so when a genuinely better model ships you switch to it and keep the same context. Same tabs, same sessions, same half-finished Notion doc. The reasoning upgrade arrives and immediately has something to reason about. Compare that to the pattern most people are stuck in, where a smarter model shows up and you celebrate by pasting slightly more text into a chat window that has never seen your inbox.

So the recursive loop compounds on the model side. The context side compounds on yours, and it compounds by being boring: staying logged in, keeping the tab open, not building a fourth integration.

What I’m still unsure about

Whether any of this holds if agents eventually get identities of their own. Maybe in three years every SaaS product issues agent credentials natively and the borrowed-session thing looks like a transitional hack. I’ve poked at that idea before and I still can’t decide. Provisioning is slow, though. Standards bodies are slower.

Either way, the model improving itself doesn’t put it inside your Workday tab. Somebody still has to log in, and for now that somebody is you, sitting right there with the session already warm.