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
Two AI models are involved — the primary agent decides, a separate auditor agent reviews — but the real safety net isn't either of them. It's a budget guard that runs as plain deterministic code: it reads the last 24 hours of confirmed spend from the ledger and vetoes the payment if it would exceed a pre-approved cap, regardless of what either model concluded. Two AI opinions can still both be wrong; arithmetic against a real ledger can't.
The same principle shows up in verification: after the payment SDK reports a transaction hash, the agent doesn't just trust that string. It independently reads the Transfer event straight off the chain and confirms the amount and recipient match before marking anything as confirmed.
What surprised us
- The x402 "exact" payment scheme isn't a simple transfer — it's a signed EIP-3009 authorization (EIP-712), and getting that signature right against a raw wallet custody API (rather than a convenience SDK that hides the details) required manually reconstructing a piece of the EIP-712 payload —
types.EIP712Domain— that higher-level libraries derive for you and never surface. Missing it would have made every real signature fail silently. - Serverless reuse is a real footgun for agent state: an agent object built once at deploy time gets reused across many unrelated sessions on the same instance. Anything session-specific has to come from the framework's own per-invocation context, never a variable captured when the agent was constructed — we found and fixed this exact bug in code review before it ever reached a user.
- The only real third-party x402 provider we found only accepted Base mainnet — proving the payment loop for real would've meant spending real USDC just to validate a demo. So we built a second data provider ourselves, speaking the same official protocol against a testnet, and used that to prove the full loop end to end for free.
What's next
More providers, adaptive routing between them, and eventually letting the agent's own spend history inform its own budget policy — bounded, always, by the same principle: two AI opinions are a signal, not a lock. The code that actually moves money stays boring, deterministic, and independently checked.
Code: github.com/ibcorreaai-oss/agentpay-procurement-agent
Demo video: youtu.be/XsQc7Wyndxc
Built for the Taskmaster category of Google's All Things Agentic Hackathon.
This post was written for the purpose of entering the All Things Agentic Hackathon.
Top comments (0)