That’s not adjoining to what AI brokers want. That is what AI brokers want. Lengthy earlier than LLMs existed, assistive applied sciences proved the core thesis: Machines can navigate interfaces, semantics can outlive presentation and construction can substitute for imaginative and prescient. Screenshot-based brokers spend tokens rediscovering information the browser already is aware of. The accessibility tree already comprises position, identify and state in structured type. In my very own agent work, switching from screenshot-based to DOM-native execution minimize per-action latency from 2–5 seconds to beneath 500ms and token price by an order of magnitude.
Accessibility proved the thesis. Now we want the following layer
Probably the most clarifying realization I had was this: Accessibility had already solved a big portion of the issue brokers face. Accessibility offers machines a approach to uncover interfaces. It exposes out there controls, their names, their states and their relationships. However discovery will not be execution. The accessibility tree can determine a button named “Checkout” and point out whether or not it’s disabled. What it can’t present is a contract for the motion itself. For instance, what inputs it accepts, what preconditions are required and what state modifications it produces.
One rising response to this hole is WebMCP, which introduces a browser-native approach to expose typed capabilities that brokers can invoke straight. When a type area has no express agent annotation, Chrome’s declarative API derives the parameter description from the related