Guide to AI Agent protocols: A2A, MCP, UCP, WebMCP...

本站添加: 2026-05-12

[00:00:00]All right, so let's say you have this multi-agent architecture within your backend. One of your agents is an orchestrator, so it decides which other agent to delegate the specific task to.

[00:00:10]And let's say we have a we are a travel website. So, one of the agents is specialized for booking hotels, and another agent is specialized for booking flights. So, whenever we get a request from a user such as, "I want to go to Japan, and my budget for a hotel is this and this, and I want to fly on this and this date." This agent runtime or this orchestrator agent is going to dispatch specific tasks to these two separate agents. Now, before I come to the point of the video, uh this sketch is from one of my previous videos where I was talking about AI era backend and how it normally would look like, where we have a normal API gateway, good old microservices, but we also have the AI gateway, which decides which agent to talk to. And now, the interesting part here's here. So, these arrows are how agents usually talk to each other. And also, this arrow tells us how agents actually talk to MCP servers, okay? The MCP server then decides which, for example, vector database to talk to or what to search on the web, and so on. Basically, it exposes uh some tools that the agent runtime is then able to use. Now, how is all of these defined? Well, there are obviously protocols, right? The P and MCP actually stands for uh word protocol, and MCP is one of the protocols. So, if we go to this website from Google, and it's called Developers' Guide to AI Agent Protocols, we have a very good overview of all the protocols that are available to us nowadays. So, obviously, the first one being MCP, and you heard about this multiple times.

[00:01:54]Long story short, MCP basically gives us a client and a server. So, if we look at this graph, basically the client in this case is going to be an agent runtime.

[00:02:04]Let me actually go to this graph, maybe it's better. So, let's say the user says, "Book me a flight to Tokyo." and we have this one agent and it has its own MCP server, okay? Then there's another agent and it has its own MCP server, but in this case, this is going to be our client and the or MCP client and this is going to be the MCP server because it exposes different actions to the client and then can perform different actions such as going to vector database or actually talking to a different agent. And if we look at this code from Google ADK, this is one of the actually frameworks that Google developed for specifically for building agents.

[00:02:48]This is what it's going to look like.

[00:02:49]So, you're defining specific service, right? And MCP was originally developed by Anthropic. Now, Google decided, "Well, we're also capable of doing something. So, why not come up with this part?" So, A2A is another protocol, so agent-to-agent protocol. Basically defines how agents, so it can be an MCP, obviously, or MCP server, but it can also be just an agent itself. How are we going to talk to agents in the well? So, this is also an A2A protocol communication, right? So, what is this actually? Well, this is something very similar to how microservices work, right? This is the agent-to-agent protocol and if you look here, basically there's a specific notation on how these agents are going to get discovered. So, if we go here and we're going to see that there's something that you need to expose, okay? And it has to live under.well-known/agent-card.json.

[00:03:54]I think this can also be just agent.json or card.json, but the best practice is agent-card.json.

[00:04:02]And you're going to be exposing something like this. You have to expose a name, a description, a skill URL, and a version. If you're a developer and if you're an experienced developer, you know that this really reminds us of the Open API notation, which probably you definitely use Swagger, so you're familiar with this, too. At least the Swagger UI where you have the Open API version, the servers, the paths, and summary and a description for every endpoint of your API backend, right? So, this is something similar.

[00:04:34]But how does the A2A protocol works? So, it's not that complicated. And by the way, it's not that mature, so you're not going to see it that much in the wild yet because it's still something that companies are trying to adopt. But basically, you have to discover what an agent has to do by querying this agentcard.json as you saw. Then the A2A server actually returns the card, and then the client, let's say if if it fetches five cards from five different A2A servers, it has to decide which A2A server to talk to. And this is something that's not actually defined within this protocol. The client has to decide by themselves, and this is pretty much open for your implementations. So, maybe a better graph. So, we basically fetch the card, the server returns us the card, and then we decide which server to talk to, which server that fits our needs best based on the name and the description, right? Very similar to how MCP servers are exposed to the MCP client. Now, you can use the ADK, but there are also a separate SDK called A2A-SDK for Python. So, you're pretty much free to choose how you want to develop this, but I would go with a Google's ADK obviously because this protocol is from Google for now. Now, there's also another or different types of protocols that I'm going to quickly tell you about so that you're aware of what is coming next. This is probably going to get adopted and going to become widespread in the following months or years. So, there's a universal commerce protocol. So, basically standardizes the shopping life cycle into module capabilities through typed request and response schema. So, this is something that has been developed by shopping companies as you can see a Shopify, Etsy, Walmart, and so on. And we scroll down here, you're going to see that there's this specific notation and this is basically the standard for AI agents and this is how they have to query and how they have to return the data. Basically, it's just another protocol. And we also have one here, agent payment protocol, and this is kind of an extension to the UCP protocol defined here. And by the way, the code is quite easy to understand. I will link it in the description for you to read, but AP2 is basically responsible for payments. And we also have agent user interface protocol as well as the agent user interaction protocol. Now, I'm not going to go very deep into those because this is still something that's in development, but there's also another one that's not in this article. It's called WebMCP. All right. WebMCP is also released, and I think it's already being used by some companies, but WebMCP would look like this. If we have, let's say, a form, you're going to see these parameters or attributes called tool parameter description, customer's phone number, and so on. Basically, you're you're adding this parameter to every form field. Now, why are we Why do we need this? Well, some AI agents go out into the internet and try to make a booking for you, right? Let's say, if you're trying to make a booking in this French bistro, your AI agent is going to go to the website, and then it has to interpret what the website is about and how to book it. And to make its life easier, you actually can add this parameter into your input fields, and then the AI agent is going to be is going to have an easier time understanding your website. Obviously, this is another protocol, so just the way to standardize. And there's an imperative and a declarative way, and the imperative way is done with with the help of JavaScript. So, let's [snorts] say if we look at the register tools, you basically register different tools within your JavaScript code, or you can just uh use attributes within your HTML like this. If you guys learned something new, and this was a quick video to tell you about what's happening within the field of protocols, and if you like this, smash like and subscribe. I'm going to see you in the next one.

[00:08:47]Goodbye.