Deploy Agents with A2A on LangSmith Deployment

Key Takeaways

A2A is a protocol released by Google in 2024 that enables agents across frameworks to communicate. It's become the standard for agent-to-agent communication. And with LangSmith deployment, you get A2A support right out of the box. Hi. I'm Harry from LangChain. In this video, we'll cover key A2A concepts, deploy simple agent with LangSmith deployment, and then show how to use its A2A capabilities. So, before jumping into the demo, I actually want to go over a couple core concepts in A2A. So, first is this agent card, which is a piece of JSON that basically describes the agent's capabilities. So, this will contain things like whether the agent supports streaming, the agent's URL, and a bunch of other meta metadata. I'll show this off later, and it'll become more clear what exactly this does. Then, there's a concept of a task. So, a task basically represents a request-response within a certain context. And this actually maps very similarly to the concept of runs in LangSmith. And last, there is this concept of a context. And a context groups a bunch of task objects together, and it becomes important in like a multi-turn conversation where I might have a bunch of tasks, but a single context for the entire conversation. And this is very similar to the concept of threads that we have. So, for the demo, I'm going to be using this test repo here, which actually comes with the basic agent with some tools, and also comes with the helper Python script that uses the official A2A Python SDK. Cool. So, I've cloned this locally, um and I've opened up the project up in WinSurf. Um I copied the env.example and set up uh .env, and I specified API key, and I'm going to be using a Anthropic for this example. And I've run a UV sync dev here also. So, if you go look at the actual agent code, I'll quickly show you what's actually happening here. So, we are setting up a basic create agent, and we have four tools, uh one that tells the time, one that just does some calculations, and then a read and send email tool, uh to mock uh an actual email tool, and we'll actually demo how uh human in the loop works with A2A with this tool as well. So, I'm going to do a quick UV run LangGraph deploy to deploy this. Nice. Okay. So, our deployment's done. I'm going to go hop over to LangSmith, go to the deployments page, and here's my deployment. So, I'm going to press this connect. You can see the exact shape of our graph here. And I'm actually going to go to the A2A tab here. So, like I mentioned before, here's the actual agent card, just a simple JSON object, and you can see um we have the URL right here for our actual agent, and you can see things like, yeah, we support streaming here. And there's a quick uh usage snippet here where we actually just show off sending a quick message. But, I'm actually going to show you how this works with Google's official A2A inspector. So, I'm going to copy this URL here. New tab. Going to go to A2A inspect.com. Going to paste in the URL here. And I'm actually going to do I'm going to have to set up API key auth here. So, I would put an API key in here. [clears throat] Once you've set that up, we can try to connect. Says our agent card is agent card is valid. This is the same thing I showed off in LangSmith's UI. And then we can actually just send a chat. So, I'm going to trigger the calculator tool here. So, I'm just going to say, "What is 3 * 2 * 2?" Send this. And this is all going through the official, you know, A2A interface. We get our answer here. And yeah, if you click on one of these, you can actually see what's being sent. So, here we can see a context was generated. A task was generated here, and you can see our actual text, and then can actually see the agent's response here as well. So, I'm going to show you how human in the loop works with the official A2A client. So, if you remember, we set up this human in the loop middleware here, and we wanted to block on this mock send email tool. And we're going to basically going to prompt the user to make some sort of approve, edit, or rejection statement. So, in the sample code, there's this test A2A HITL human in the loop um file here, and and here we are actually going to use the official A2A client. Um I've pasted in my agent URL here that I got from uh the LangSmith UI. And we set up a client here. So, we're going to send uh a message first asking to actually send an email. And then, after that, you can see what we do here is in order to approve the interrupt, we actually have to send this data part back. And in this data part, we actually send this resume payload. And we're going to say type approve, and [snorts] this will actually approve the task. Remember, this is all going through the A2A client. So, I'm going to quickly run this script. Cool. We can quickly s- scan the output here. So, first we say that we wanted to send the email, and you can see that in our response, we actually got this tool execution requires approval. And we need to either approve, edit, or reject this. And you can see the task is actually in this input required state, which is part of the official A2A spec for for task task states. And then, when we actually resume the interrupt, we get this done. Your email has successfully been sent. Cool. So, yeah, that's the A2A integration with uh LangSmith deployment.