Claude Can Now Click ref_2 Instead of Guessing Pixel 640, 380
The browser use tool shipped with 31 member tools and element references read straight off the accessibility tree. The interesting part isn't that Claude can see the page — it's that the reference expires.
NeuroX AI · August 25, 2026
Anthropic launched the browser use tool on 19 August — browser_toolset_20260801, 31 member tools enabled by default, built for agents that live inside webpages instead of on a whole desktop.
The capability that matters is easy to skim past. read_page returns the accessibility tree as text, with a handle on every element:
link "Getting started" [ref_2]
textbox "Search docs" [ref_3]
Claude clicks ref_2. It does not find the link in a screenshot and estimate a coordinate first. The docs are explicit that computer use works "through screenshots and coordinates alone" — and we've written about what that costs: 83.5% on benchmarks, 20.6% on long-horizon real work. A meaningful slice of that gap is pixel arithmetic against a rendered image.
Now read the constraint printed next to the feature. References "stay valid until that tab navigates or its DOM changes materially." A ref is a lease, not an address. Any app that re-renders between the read and the click invalidates the handle — and because a multi-call turn is a batch that stops at the first failure, every later action returns Not executed: an earlier action in this turn failed.
That's the right design, and it's the thing to engineer around. read_page caps at 50,000 characters. find returns at most 20 matches. Every handle either of them hands back is stale the moment the page moves. The agent's job is not to remember the page. It's to re-read it before it acts.