Sign in to continue learning
A2A: The Agent2Agent Protocol
Intermediate
1h27m
Topics
Agents
We mentioned earlier in the course how A2A and MCP are complementary protocols. So let's actually build an agent that uses both. For this example, you'll create a healthcare provider recommendation agent using LangGraph and OpenAI GPT OSS on Vertex AI. You'll build an MCP server that provides a list of fictional doctors with their locations and specialties. Then you'll connect it to the agent, which you'll finally expose as an A2A server. Let's go. Now, you're going to create an A2A agent which also uses MCP. And this agent is being built using the popular AI agent framework LangGraph. So first, we need to create an MCP server. We're going to be creating this MCP server using the FastMCP library. So I'll again use the writefile Jupyter magic to define this as a separate file called mcpserver.py Once again, we'll import some libraries including FastMCP. Then I'll create the FastMCP server, which I'll call doctorserver. This agent is going to be searching for healthcare providers using a specific MCP tool. I'll load in a JSON file that I have stored in the data directory. And I'll load it in as a list called doctors. Then I'll use the MCP tool decorator to make this next method we're going to create into an MCP tool. This method is called list_doctors. So I'll have it take in a US state name and a city. And it'll return a list of dictionaries. And remember to put in a detailed docstring, because this will be used by the LLM to decide which functions to call and when. This method will first do some input validation to make sure at least a city or a state is provided. We'll normalize the city and state to do case-insensitive matching. So now, we'll filter down the list by state and city. using a list comprehension. FastMCP will automatically convert this list of dictionaries into JSON when sending to the model. Now, I'll kick off the server with dunder main and run the MCP server over standard IO. This means the server will be running locally. Now that you have the MCP server set up, you'll now build the health provider agent using LangChain and LangGraph. First, I'll import all the libraries, including LangChain. I'll authenticate to Google Cloud. In order to use the LangChain OpenAI connector with Vertex AI, you'll need to override the default URL. So I'll use this custom URL with the Vertex AI OpenAI endpoint. I'll use nest_asyncio to enable async code to run in a Jupyter environment. Now I'll define the MultiServerMCPClient for LangGraph. I'll call it find_healthcare_providers, and run the MCP server over standard IO, using uv run mcpserver.py. Just to clarify in this case, the MCP server is running as a subprocess of the Python script, not as a separate HTTP server. Then, I'll convert the MCP client into a LangGraph tool. Now, I'll create the agent using the LangChain create_agent method. I'll use the ChatOpenAI class for the model. The name will be OpenAI GPT OSS 20 billion model as a service. I'll add the Google Cloud credentials as the API key. and I'll use the Vertex AI URL as the base_url. I'll add in the tools, give this agent a name, and add in the system_prompt, which will task the agent with using the MCP tool to find healthcare providers based on the user's query. Now, in my experimentation, this model likes to try and include real doctors instead of the fake ones in the tool. So I need to prompt it to just return providers from the tool. Now, I'll test this agent with a user prompt. In this case, I'll ask this agent for any psychiatrists in Austin, Texas. I'll print out the last message as markdown. And it looks like Dr. Coffey is available in Austin for psychiatry. So, the LangGraph agent works. Now, let's make it A2A compliant. LangGraph serving, LangSmith, has a built-in A2A integration, but it works very differently than the other integrations in this lesson. So for consistency, I'll follow a similar process to the first agent created using the A2A SDK. This is also what you would do for any agent framework which doesn't yet have an A2A integration. We're going to again use the writefile Jupyter magic to add this to agents.py. In this case, we'll be appending it to the end because we already have the Claude agent from earlier in agents.py First, we'll import some libraries, including everything from LangChain and our helpers. And I'll create a container class for the LangGraph agent called ProviderAgent to make it easier to abstract away the details. We have an init method to initialize the agent and set all the important fields. Then we also have an initialize method. Because the MCP client connection is asynchronous, you need to use a helper initialize method to avoid race conditions when initializing. This is where we actually create the agent. Then we'll put all of the agent communication logic within the answer_query method. It'll take in the prompt and it'll send out a response as a string. Now, I'll test this class we just created to make sure it works as we expect. Okay? And it's working just like we expect. Now, I'll wrap up this provider agent in an A2A agent executor. We'll save it as a file called a2a_provider_agent.py We'll load in the libraries, create our executor class. In this case, I'll do a lazy initialization of the agent. In the execute method, I'll make sure that the agent is initialized, and then I'll get the user input just like before, pass the prompt into the answer_query method. Then I'll enqueue the agent message to the event_queue. And I'll pass on the cancel method just like before. Now, let's look at the main method. I'll load in environment variables, define the HOST and PORT, create the AgentSkill, which will basically just be describing the MCP tool. Then load up the AgentCard with all the details. I'll create a request handler and an A2A Starlette server and run it with uvicorn, and we'll have dunder main to actually run all of this code. Now, I'll run it in terminal 3. It looks like it's up and running. Now, in the next lesson, we'll connect to this agent with another example of an A2A client using Microsoft agent framework. Then after that, we'll combine this agent
