Four minutes of live runs: giving the agent better "eyes" beats giving it a better model
Browser Agents Don't Need Better Models. They Need Better Eyes. - Kushan Raj, ARK · Kushan Raj
4 min total·Actually worth watching closely: ~3 min·3 must-watch clips
- 0:00 – 0:18Listen
A great idea nobody actually uses
Opens with the state of play: browser agents are such a cool idea, but the adoption hasn't happened — the speaker barely uses them himself. The goal he sets is making them faster, cheaper, and more reliable.
The question isn't whether it can be done; it's that being slow and unreliable keeps them out of daily use.
Purely spoken background and motivation, with nothing on screen you need to watch. Fine to have on while doing something else.▶ Jump to 0:00Speaker · Kushan Raj - 0:18 – 0:52Watch
Just how slow today's agents are
Demos current agents on a browser benchmark that requires long sequencing of tasks: 10 to 20 seconds just to click the start button, and a 30-step task already struggling at step one.
Failure on long sequences isn't an occasional glitch — every single step bleeds time, and it compounds into inevitable collapse.
You only feel the slowness by watching the seconds tick by; put into words it's completely lost. Keep your eyes on the screen here.▶ Jump to 0:18Speaker · Kushan Raj - 0:52 – 1:36Watch
The counterintuitive call: the model is fine
The core hypothesis: models are pretty smart, but it's the infra around them that sucks. In the failing runs the agent isn't unable to think — it keeps trying to click and can't understand what changed on the page. This is also the first appearance of his own build, running the same site much faster on a much cheaper model.
Blaming "the agent can't do this" on model capability points you in the wrong direction; what's missing is any sense of page state.
The argument and the first demo of his own agent are layered together — the faster run on screen is the evidence itself, so audio alone loses the comparison.▶ Jump to 0:52Speaker · Kushan Raj - 1:36 – 2:14Watch
Live run one: downloading an ID document
The same task — downloading an Aadhaar — with two agents run side by side. The screenshot-loop Claude agent clicks the first button, then takes a screenshot, scrolls for some reason, takes another screenshot; the whole process takes two minutes. The speaker's agent, on a much cheaper model, is done almost instantly.
A screenshot shows only one particular snippet of the page, so the agent has to scroll around feeling its way — that's where the time goes.
The side-by-side gap is the most vivid moment of the whole talk; one spinning its wheels while the other has already finished is something you have to see.▶ Jump to 1:36Speaker · Kushan Raj - 2:14 – 2:51Watch
Live run two: picking a date on an unfamiliar site
Switches to a Canadian trek booking site with an interface that's unfriendly to agents. Claude never manages to pick a date and just gets stuck; the speaker's agent selects the date and completes the booking.
What really separates the two is unfamiliar, non-standard sites — controls like date pickers expose the weakness of the representation layer fastest.
The stall happens in those few seconds on the date control; only by watching the cursor probe around the calendar do you see which step it actually loses on.▶ Jump to 2:14Speaker · Kushan Raj - 2:51 – 3:24Skim
How he plans to ship it: open source and product shape
The speaker explains the code isn't super defensible, so he's thinking about open sourcing the project. The product shape is still open: maybe an API that takes your URL and your intent, executes it for you and gives the result back, or a website, or a plugin.
The moat isn't in the code, so he'd rather open source it for adoption — which also explains why the whole talk is about infra rather than models.
Mostly plans narrated over a few command examples — directional content, so glance at the screen and take the conclusion.▶ Jump to 2:51Speaker · Kushan Raj - 3:24 – 4:21Listen
The mechanics: the whole page in 1,800 tokens
The key numbers: the full DOM would be around 20,000 tokens, a screenshot covering only one particular snippet about 1,100, and his markdown compressed representation about 1,800 — and with it the model sees the entire website. Alongside it runs end-to-end page tracking that tells the agent what has popped up, what is now gone, whether the thing blocking the target has been removed, and why a click didn't take.
The value isn't just saving tokens — it's seeing the whole page plus knowing what just changed, which is what makes planning long sequences possible at all.
This closing stretch is essentially the speaker reciting numbers and mechanisms with nothing more being demoed; just listen for the token counts and the feedback items.▶ Jump to 3:24Speaker · Kushan Raj