A recent Sentience blog post pits a cloud‑based GLM‑4.6 baseline against a fully local stack for the classic Amazon “search → first product → add to cart → checkout” flow. Four runs show the cloud demo succeeded in one pass, while the local configuration completed every step across seven re‑runs.

The local stack isolates three components: a DOM‑pruning pass that compresses the page to a semantic snapshot, a DeepSeek R1 planner that emits step intents, and a Qwen ~3B executor that issues concrete CLICK or TYPE commands. After each action, Sentience’s runtime asserts URL changes or element presence, retrying only on explicit failures.

Results prove that verification—inline assertions over structured snapshots—delivers reliability where larger vision models often stumble. Token usage dropped 43 % compared with the cloud run, and no API fees accrued because inference stayed on a Mac Studio. Developers may now consider verification layers the primary lever for robust browser agents.