Accessibility Tree Cuts AI Agent Latency Below 500ms, Slashing Token Costs 10-Fold
Updated
Updated · InfoWorld · Jul 7
Accessibility Tree Cuts AI Agent Latency Below 500ms, Slashing Token Costs 10-Fold
3 articles · Updated · InfoWorld · Jul 7
Summary
DOM-native browser agents using the accessibility tree cut per-action latency from 2–5 seconds to under 500 milliseconds and reduced token costs by an order of magnitude, the expert said.
The gain comes from using the web’s existing machine-readable layer—roles, labels, states and relationships—instead of screenshot-based systems that spend tokens rediscovering what browsers already know.
That semantic layer still stops short of execution: it can identify a "Checkout" button and whether it is disabled, but not the action’s inputs, preconditions or resulting state changes.
WebMCP is presented as the next step, exposing typed browser capabilities and deriving parameter descriptions from labels or aria-description, while poor frontend practices like unlabeled div-based controls and stale ARIA states leave agents blind.
The broader takeaway is that accessibility work now doubles as agent infrastructure, with stable identifiers, synchronized ARIA state and explicit action metadata forming the basis for reliable web automation.