Building an AI Agent That Shops, Gets Audited, and Pays — On Its Own

Written for the All Things Agentic Hackathon (Google).

Most "autonomous agent" demos stop at reasoning. This one doesn't: it decides it needs a piece of data, compares real prices across multiple third-party providers using the x402 HTTP payment protocol, has its spending decision reviewed by a second AI agent, and then actually pays — real USDC, signed via MPC custody, verified independently on-chain.

The shape of the problem

Payment is the part everyone skips in agent demos, for a good reason: it's the part where a hallucination costs real money. So the design question wasn't "can an agent pay for something" — it's "how do you build an agent that can be trusted to pay for something, autonomously, repeatedly, without a human approving every transaction."

What actually enforces trust here