Researchers at Sentience ran four Amazon shopping flows to test whether a small local model could match a cloud‑based giant. By gating every browser action with a verification layer, they proved that a ~3 B‑parameter executor can finish the search‑to‑checkout sequence without any vision model today.

The architecture splits reasoning from action: DeepSeek R1 plans intent, while Qwen selects DOM clicks and types. A WASM pass trims the DOM to a structured snapshot, dropping roughly 95 % of nodes, and a Jest‑style assertion verifies each step before proceeding for reliable automation every run.

Results show local autonomy achieving 7/7 steps with 11,114 tokens, while the cloud baseline used 19,956 tokens and 60 s. The study highlights that verification drives reliability, not sheer model size, and that token efficiency stems from interface design rather than larger models in practice today.

Industry teams can adopt this pattern to build robust browser agents without costly GPU clusters. Future work may extend the verification loop to multi‑site commerce and integrate reinforcement signals. The approach signals a shift toward structured, assertion‑driven automation in web‑interaction AI today and ahead of.