Letting agents operate web pages starts with controlling how much they observe at a time

LangChain posted a short piece on X about browser use, pointing out an easily overlooked cost: having an agent operate websites designed for humans can quickly fill up the context window.

Browser use sounds straightforward: the agent opens a page, reads content, clicks buttons, and fills out forms. The trouble is that pages themselves are heavy. The DOM, accessibility tree, screenshots, text snapshots—any one observation method can consume a large number of tokens. The workbench is only so big; the more of the page you spread out, the less room is left for reasoning and memory.

LangChain says it selected and tuned browser tools in Managed Deep Agents, using Browserbase's Stagehand under the hood. The original post is here: LangChain's original post. Sydney Runkle wrote a longer article: article link.

There are two concrete approaches in the replies. One commenter said to first export the page as a cropped snapshot before letting the agent act; it isn't complicated. Another commenter added that page dumps can quickly eat up the context. Both points lead to the same thing: the observation surface needs to be narrowed. For developers, this looks more like an observation budget problem than a model parameter problem. For ordinary users, it can be understood as not spreading the entire manual out on the table at once, but only turning to the page you need right now.

There's another detail from Alek: deferred form validation. The page goes quiet, but Submit is still disabled; whether the agent can continue depends on whether the next action is available. Kala also mentioned that those 8-second naps add up, and traces caught them. The browser doesn't get lighter just because an agent shows up. In tool design, how long to wait, how much to observe, and when something counts as ready are all hard problems.

What makes this worth reading is that it puts the selection and tuning of browser tools on the table. Being able to click is only the starting point; how much of the page to observe each time and how to determine when a page is ready are what determine whether an agent can remain stable on real websites.