For thirty years, commerce sites have treated automated buyers as the enemy. Now customers are sending AI agents to shop on their behalf, and the card networks, Google, OpenAI, and Stripe have all shipped protocols that assume merchants will let them in. Giving an agent tools turns out to be the easy part. Giving it bounded authority to spend someone else's money safely is the real engineering problem.

This session walks the merchant-side "front door" one step at a time: discovery, connectivity, identity, authorization, approval, payment, and evidence. Each step is motivated by a live failure against MockHub, a working mock ticket marketplace: an identity binding that lets an agent read the wrong customer's orders, a spending limit that checks the wrong number, and an agent that approves its own purchase because the approval tool was reachable. Then you see the correction as real code, with Spring AI and the Java MCP SDK doing the work.

You'll leave with a seven-step checklist, an understanding of where MCP, AP2, ACP, UCP, and TAP each fit, and a public repository with runnable examples for every step.

- Format: presentation with live demos against a hosted platform. Recorded fallbacks exist for every demo if the venue network fails.
- Duration: 50 to 60 minutes. A half-day or full-day hands-on workshop version exists, with labs in Java, Python, and TypeScript, if the organizers want one.
- Audience: intermediate. Developers and architects who own a transactional system, plus tech leads and product people deciding whether to open it to agents. No prior MCP experience needed.
- Requirements: projector and internet. Nothing for attendees to install.
- Materials are public at github.com/kousen/agentic-commerce.
- The specification landscape changes quickly. The talk is updated before every delivery to reflect current status.