The Shipping and Tax Problem: Why Agents Struggle to Close the Sale

Rye solved shipping and tax by solving automated checkout end-to-end, with an agent that can successfully and repeatably navigate the long tail of e-commerce sites.

The system runs in two phases. On the first pass, an LLM navigates a merchant’s checkout flow in exploratory mode. A filter simplifies the DOM by suppressing popups, banners, and irrelevant fields so the model can focus on what matters. Guardrails catch obvious errors. This exploratory run produces a trace of how to reach the point in checkout where the merchant reveals shipping options and tax.

On the second pass, Rye converts that trace into a deterministic workflow. On subsequent orders, the workflow executes step by step without invoking the LLM. If the site changes, only the failing step is rerun with an LLM, after which the workflow self-heals. Because the LLM is limited to exploration and occasional healing, Rye can rely on smaller, faster models. Latency is no longer dictated by expensive reasoning cycles.

None of these techniques would matter if we couldn’t reliably get to the checkout page. Rye’s Merchant Risk Adapter for fraud-safe navigation leverages residential IPs, geo-proximal sessions, and human-like pacing to appear as a real shopper to merchants.