WEBVTT

00:00:01.290 --> 00:00:02.116
Hi, I'm Caramel!

00:00:02.116 --> 00:00:04.044
Jev can only pick from a list.

00:00:04.044 --> 00:00:05.696
It can't write a single word.

00:00:05.696 --> 00:00:08.174
Paired with a writer, it tops a web-agent board.

00:00:08.174 --> 00:00:08.725
Here's how.

00:00:10.350 --> 00:00:12.673
The idea: split the agent in two.

00:00:12.673 --> 00:00:17.320
A decision model picks the tool, and a small writer fills in the arguments.

00:00:17.320 --> 00:00:19.643
Jev never types, and Mercury never chooses.

00:00:21.508 --> 00:00:24.589
Under WebMCP, the page registers its own tools.

00:00:24.589 --> 00:00:28.054
Those tools, plus finish and abstain, become the menu.

00:00:28.474 --> 00:00:30.815
Jev reads the state and picks one option.

00:00:30.815 --> 00:00:34.033
That's all it can do: choose, with a probability for each.

00:00:34.453 --> 00:00:38.443
Mercury, a fast diffusion model, writes the arguments, forced into the

00:00:38.443 --> 00:00:39.894
chosen tool's own schema.

00:00:40.314 --> 00:00:43.971
If an argument is wrong, a well-built tool replies with the legal

00:00:43.971 --> 00:00:46.104
values, and the next step uses one.

00:00:47.900 --> 00:00:50.976
Driving the page, every click is a separate decision.

00:00:50.976 --> 00:00:53.711
A checkout takes a median of twenty-two actions.

00:00:54.131 --> 00:00:56.626
With tools, one call can hold a whole journey.

00:00:56.626 --> 00:00:58.012
The same checkout takes four.

00:01:01.192 --> 00:01:06.121
The headline: two hundred and forty-five times cheaper than GPT-6 Astra on screenshots.

00:01:06.121 --> 00:01:08.396
But two things changed at once.

00:01:08.816 --> 00:01:13.044
About fifteen of it is the interface: Astra itself, moved to

00:01:13.044 --> 00:01:14.582
WebMCP, mostly reading less.

00:01:15.002 --> 00:01:18.672
About sixteen is price: the same interface, with a pair that costs

00:01:18.672 --> 00:01:19.895
far less per token.

00:01:21.308 --> 00:01:25.675
Most of the margin is the interface, available to any model that uses it.

00:01:25.675 --> 00:01:28.170
The decision model wins on price per token.

00:01:30.608 --> 00:01:34.814
To recap: one model picks, one types, tools compress the journey, and

00:01:34.814 --> 00:01:37.267
the big multiplier is two smaller ones.

00:01:37.687 --> 00:01:39.771
Every source is in the full article.

00:01:39.771 --> 00:01:40.367
I'm Caramel.

00:01:40.367 --> 00:01:40.665
Bye!
