Agentic Checkout

Agentic Checkout ("Buy for Me") is an AI agent that completes multi-step purchase workflows end to end. I took it from a 0→1 idea to its public debut at Google I/O; it was recognized with the 2026 Webby Award for Best AI Agent.

How it started

The idea came first, from my own pain point: the thing I wanted would drop in price — and I still had to watch it and buy it myself. I imagined the product that should exist — an agent that completes the purchase you've already decided to make — and only then went looking for where it could live. Upgrading the existing price-tracking product was that path: it could tell millions of shoppers when to buy, but carried them no further. I built the first prototype with a small team, winning first place at an internal hackathon.

The unglamorous middle

Turning a winning demo into a real product meant evaluating build-vs-reuse paths across existing checkout infrastructure and AI web agents, designing the system architecture, and aligning shopping, payments, and AI teams around one MVP scope.

The make-or-break: reliability

Checkout is adversarial and high-stakes — an agent that usually works is a demo, not a product. Through systematic agent design, prompt iteration, and eval pipelines, checkout success went from roughly one-in-four to production-grade on target flows. That number is what turned demos into an official product track — and it's where I learned the lesson that shapes all my work since: an agent is only as good as the system that verifies it.