Developer’s Guide to AI Agent Protocols Google Developers Blog

Izzo If you've ever tried building an AI agent that actually does useful work, you've hit the integration wall.

Izzo Welcome back to Exploring Next, I'm Izzo. This is episode two-thirty-five, and Boone, Google just dropped something that could fix agent development's biggest pain point.

Boone The acronym soup, right? MCP, A2A, UCP — I was ready to roll my eyes until I saw what they're actually solving.

Izzo Exactly. Every agent needs to talk to databases, other agents, payment systems. Right now that means custom integration code for literally everything.

Boone And the maintenance nightmare when any of those services change their API. I've been there.

Izzo So Google's Agent Development Kit introduces these three protocols that standardize how agents connect to... well, everything. Let's build something real.

Boone They use a restaurant supply chain agent as the demo. Kitchen manager that needs to check inventory, get supplier quotes, place orders, handle payments.

Izzo Perfect use case. Boone, walk me through Model Context Protocol first — this is the data connection layer?

Boone MCP eliminates what I call integration debt. Instead of writing custom tools for every API endpoint, you connect to MCP servers that advertise their capabilities automatically.

Izzo So the PostgreSQL team maintains the database MCP server, Notion maintains theirs, Mailgun maintains email — I'm not writing or updating any of that code?

Boone Exactly. Look at this kitchen agent setup — McpToolset connects to three different services with identical patterns. Database queries, recipe lookups, email sending.

Izzo The import structure is clean. You're just pointing to server URLs and passing credentials.

Boone Right, and when Notion adds new capabilities to their MCP server, your agent gets them automatically. No code changes, no redeployment.

Izzo That's huge for product velocity. But what about when your agent needs expertise it doesn't have? Like wholesale pricing or supplier ratings?

Boone That's where Agent2Agent protocol gets interesting. Each remote agent publishes an Agent Card at a well-known URL — slash dot well known slash agent card dot json.

Izzo Hold on, break down this discovery mechanism for me.

Boone Think DNS for agents. The card describes what the agent does, what skills it has, where to reach it. Your agent fetches these cards and routes queries to the right specialist.

Izzo So my kitchen agent discovers a pricing specialist, a quality grading agent, a delivery logistics agent — all through the same protocol?

Boone Exactly. And when you need to add a new supplier agent, you just add a URL. No custom integration, no API documentation to parse.

Izzo I'm giving this architecture a solid A-minus. The discoverability is elegant, but what about the commerce piece?

Boone Universal Commerce Protocol tackles checkout hell. Every supplier has different APIs — different schemas, different auth, different webhooks.

Izzo Different everything. I've seen teams spend months integrating payment flows that should take days.

Boone UCP standardizes the shopping lifecycle with strongly typed schemas. Same checkout pattern whether you're connecting via REST, MCP, A2A, or even browser automation.

Izzo Show me what that looks like in practice.

Boone The supplier publishes their UCP profile at the same well-known URL pattern. Your agent builds a CheckoutCreateRequest with line items, quantities, payment details.

Izzo And this works across different payment processors, different inventory systems?

Boone That's the promise. UCP handles the abstraction layer. Whether it's Stripe, PayPal, or some proprietary wholesale system, your agent sees the same interface.

Izzo Okay but here's my product question — who adopts this first? The agent builders or the service providers? Classic chicken-and-egg. But Google's pushing this through ADK, and they're showing real code that works today. The restaurant demo is smart positioning. Supply chain is where agents provide obvious ROI — inventory optimization, bulk purchasing, vendor management. And the technical implementation is genuinely impressive. They start with a hallucinating LLM, then add proto