- A browser agent that hit an access-denied page learned from its own failed run which URL format works and which paths trigger the block.
- The talk also described WebMCP, a way for sites to offer agents an official door, though a speaker said adoption is still low.
A browser agent that a website blocked found a way back in. It did so by studying how the site's web addresses work. Shubhankar Srivastava of Browserbase showed this on the AI Engineer show, using his AutoBrowse loop. He also covered WebMCP, a way for sites to offer agents an official door.
What was said
Srivastava's complaint is that browser agents start every run from zero. His loop adds memory. A doer operates the browser. A teacher reads the screenshots and results. It then writes lessons into a file for the next attempt.
The main demo used OpenTable. The first run hit an access-denied page. A speaker on the show said of that run: "It realized that the block is environmental because the website doesn't want this agent to do something."
By the second run, the agent had found a working address format. A query parameter with latitude, longitude and a metro ID loads a usable search page. Its notes also named two page paths where the blocks kick in, and said to avoid them. On screen, the cost fell from $6.22 to $0.51. The time fell from 189 seconds to 27.
A speaker added that such loops can go further. They said: "sometimes these research loops are so smart, they realize that browsers are not needed at all, and they just reverse-engineered the API".
Why it matters
Our reading: the demo shows a block can double as a clue. The agent learned which paths trigger the block. It also learned which address format avoids it, and saved that for reuse. A site that relies only on blocks may be showing agents where its limits are.
The alternative is to state what an agent may do. In the talk, WebMCP was described as making a website programmable. A site can say how to book a reservation. It was compared to an API that lives in the browser. Teams that run browser agents should note one more point. Learned skills rest on sites that, a speaker said, change every day.
The other side
WebMCP is early. A speaker said they have not seen much adoption yet. Their own research loop does not support it. They thought adding it would not be hard.
There is also a commercial angle. The presenter said they deploy these agents from a sales perspective. Browserbase has shipped WebMCP support and tools for site owners. The speaker also said results vary by task.
The parts of the talk we reviewed leave questions open. They do not show how OpenTable responded. They do not show whether the method beats stronger defences, or whether sites permit it. Nor do they show that an official door would stop agents from taking other routes.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Hill-Climbing Skills: Improve Agents Without Changing the Model — Shubhankar Srivastava, Browserbase (2026-10-11)
