NOTE / SAMPLE /
Browser agents should not reason about deterministic navigation
A model is useful when a step is ambiguous. Most interface navigation is not.
The expensive confusion
A screenshot can show a person where a button appears. It is a poor machine interface when the browser already exposes the document, accessibility tree, URL, network activity, and element state.
Letting a model rediscover those facts on every step adds latency and probabilistic failure without adding useful judgment.
A better boundary
Use deterministic browser control for navigation, field entry, state checks, and replay. Invoke a model only where the task contains genuine ambiguity: classifying an unfamiliar page, interpreting open-ended language, or selecting among choices that cannot be expressed as rules.
The useful architecture is not agent or no agent. It is a visible boundary between facts the software can know and decisions that deserve reasoning.
The test
If the same screen state should always produce the same action, the action probably belongs in ordinary software. If the decision requires context that cannot be enumerated economically, a model may earn its place.