The prevailing vision for AI and the web goes something like this: agents will open a browser, click around like a person, and complete tasks end-to-end. Book the flights, file the expense report, find the right form, buy the thing. Everyone gets an executive assistant. The web becomes AI infrastructure, never seen by humans.
I've worked in automation for 20 years, and something about this doesn't sit right with me.
The models will be technically capable soon, no doubt about that. But a task best suited to full delegation must be easy to write out in text and have its requirements known upfront.
How often is that actually the case?
It’s true we waste hours each day on the web's shortcomings. Retyping the same preferences. Clicking through filters we've set a dozen times before. Manually reviewing results because search couldn't express what we actually wanted. Hunting for the right spot on some awful site we've not used before.
But think about how most tasks actually unfold. You start a search with a rough idea. You see results, refine your criteria. You discover an option you didn't know existed. You compare, reconsider, change your mind. The requirements aren't front-loaded — they're distributed throughout. And the value isn't all at the end either. Half the point is the judgment calls along the way.
Full delegation assumes you can specify the destination before you've seen the map.
The anticipatory web
I love anticipatory design. When you skip back on Apple TV, subtitles turn on temporarily, anticipating that you missed something. When a service texts you a login code, iOS surfaces it right on the keyboard. These systems understand intent and remove a step.
This is the model we think about: not an agent that replaces you, but an environment that anticipates what you need and clears the path, leaving you free to learn and discover.
What we're building
We’ve been exploring what the future of the web will look like, and our instinct is that it’s more like tedium reduction, not full delegation. A human-in-the-loop, intent-aware experience that skips the repetitive mechanics of the web while preserving discovery.
That means applying the context of your preferences, what you're currently doing, and a deep understanding of the site you're on. It means building things that don't exist yet: functional maps of sites that go beyond content indexing, ways to get pages into the right state without dozens of clicks, and intelligence that works behind logins and paywalls.
Over the next few weeks we'll be sharing previews of this work, and its impact on state-of-the-art browsing benchmarks. If you're building in this space, or just tired of burning hours on the web, we'd love to have you along.