Multi-Agent Communication with the A2A Python SDK

If below a rock and you’re employed with AI, you’ve most likely heard about Agent2Agent (A2A) Protocol, “an open customary designed to allow communication and collaboration between AI Agents”. It’s nonetheless fairly new, however it’s already getting a variety of buzz. Because it performs so properly with MCP (which appears prefer it’s changing into the business’s customary), A2A is shaping as much as be the go-to customary for multi-agent communication within the business.

When Google first dropped the Protocol Specification, my first response was mainly: “Okay, cool… however what am I purported to do with this?” Fortunately, this week they launched the official Python SDK for the protocol, so now it lastly speaks a language I perceive.

On this article we’re going to dive into how the protocol truly units up communication between brokers and shoppers. Spoiler: it’s all in a task-oriented method. To make issues much less summary, let’s construct slightly toy instance collectively.

Communication between the Occasion Detector Agent and an A2A Consumer

In our programs we now have an Occasion Detector AI Agent (chargeable for detecting occasions) and an Alert AI Agent (chargeable for alerting the consumer of the occasions). Since I’m specializing in the A2A protocol right here, each brokers are mocked as easy Python strategies that return strings. However in actual life you’ll be able to construct your brokers with any framework you want (LangGraph, Google ADK, CrewAI and so forth).

We’ve three characters in our system, the consumer, the Occasion Agent and the Alert Agent. All of them talk utilizing Messages. A Message represents a single flip of communication within the A2A protocol. We wrap the brokers into A2A Servers. The servers expose an HTTP endpoint implementing the protocol. Every A2A Server has Occasion queues that act as a buffer between the agent’s asynchronous execution and the server’s response dealing with.

The A2A Consumer initiates communication, and if two brokers want to speak an A2A Server can even play the function of a A2A Consumer. The diagram beneath exhibits how a Consumer and a Server talk inside the protocol.

The EventQueue shops Messages, Duties, TaskStatusUpdateEvent, TaskArtifactUpdateEvent, A2AError and JSONRPCError objects. The Job is perhaps crucial object to grasp the way to construct multi-agent programs with A2A. In line with the A2A documentation:

• When a shopper sends a message to an agent, the agent may decide that fulfilling the request requires a stateful activity to be accomplished (e.g., “generate a report,” “e book a flight,” “reply a query”).
• Every activity has a novel ID outlined by the agent and progresses by an outlined lifecycle (e.g., submitted, working, input-required, accomplished, failed).
• Duties are stateful and might contain a number of exchanges (messages) between the shopper and the server.

Consider a Job as one thing in your multi-agent system that has a clear and distinctive purpose. We’ve two duties in our system:

1. Detect an occasion
2. Alert the consumer

Every Agent does its personal factor (activity). Let’s construct the A2A Server for the Occasion Agent so issues grow to be extra tangible.

Constructing the A2A Server for the Occasion Agent

First up: the Agent Card. The Agent Card is JSON doc used to get to know different brokers obtainable. :

• Server’s identification
• Capabilities
• Expertise
• Service endpoint
• URL
• How shoppers ought to authenticate and work together with the agent

Let’s first outline the Agent Card for the Occasion Detector AI Agent (I’ve outlined the abilities primarily based on this example from Google):

agent_card = AgentCard(  
    identify='Occasion Detection Agent',  
    description='Detects related occasions and alerts the consumer',  
    url='http://localhost:10008/',  
    model='1.0.0',  
    defaultInputModes=['text'],  
    defaultOutputModes=['text'],  
    capabilities=AgentCapabilities(streaming=False),  
    authentication={ "schemes": ["basic"] },  
    expertise=[  
        AgentSkill(  
            id='detect_events',  
            name='Detect Events',  
            description='Detects events and alert the user',  
            tags=['event'],  
        ),  
    ],  
)

You may study extra in regards to the Agent Card object construction right here: https://google.github.io/A2A/specification/#55-agentcard-object-structure

The agent itself will truly be a Uvicorn server, so let’s construct the major() methodology to get it up and operating. All of the request will probably be dealt with by the DefaultRequestHandler of the a2a-python SDK. The handler wants a TaskStore to retailer the Duties and an AgentExecutor which has the implementation of the core logic of the agent (we’ll construct the EventAgentExecutor in a minute).

The final element of the major() methodology is the A2AStarletteApplication, which is the Starlette software that implements the A2A protocol server endpoints. We have to present the Agent Card and the DefaultRequestHandler to initialize it. Now the final step is to run the app utilizing uvicorn. Right here is the complete code of the major() methodology:

import click on  
import uvicorn  
from a2a.sorts import (  
    AgentCard, AgentCapabilities, AgentSkill
) 
from a2a.server.request_handlers import DefaultRequestHandler  
from a2a.server.duties import InMemoryTaskStore  
from a2a.server.apps import A2AStarletteApplication 
 
@click on.command()  
@click on.possibility('--host', default='localhost')  
@click on.possibility('--port', default=10008)  
def major(host: str, port: int):  
    agent_executor = EventAgentExecutor()

    agent_card = AgentCard(  
        identify='Occasion Detection Agent',  
        description='Detects related occasions and alerts the consumer',  
        url='http://localhost:10008/',  
        model='1.0.0',  
        defaultInputModes=['text'],  
        defaultOutputModes=['text'],  
        capabilities=AgentCapabilities(streaming=False),  
        authentication={ "schemes": ["basic"] },  
        expertise=[              AgentSkill(                  id='detect_events',                  name='Detect Events',                  description='Detects events and alert the user',                  tags=['event'],  
            ),  
        ],  
    )
      
    request_handler = DefaultRequestHandler(  
        agent_executor=agent_executor,  
        task_store=InMemoryTaskStore()  
    ) 
 
    a2a_app = A2AStarletteApplication(  
        agent_card=agent_card,  
        http_handler=request_handler  
    )  

    uvicorn.run(a2a_app.construct(), host=host, port=port)

Creating the EventAgentExecutor

Now it’s time to construct the core of our agent and eventually see the way to use the Duties to make the brokers work together with one another. The EventAgentExecutor class inherits from AgentExecutor interface and thus we have to implement the execute() and the cancel() strategies. Each take a RequestContext and an EventQueue object as parameters. The RequestContext holds details about the present request being processed by the server and the EventQueue acts as a buffer between the agent’s asynchronous execution and the server’s response dealing with.

Our agent will simply verify if the string “occasion” is within the message the consumer have despatched (KISS ✨). If the “occasion” is there then we must always name the Alert Agent. We’ll do this by sending a Message to this different Alert agent. That is the Direct Configuration technique, that means we’ll configure the agent with a URL to fetch the Agent Card of the Alert Agent. To do this our Occasion Agent will act like a A2A Consumer. 

Let’s construct the Executor step-by-step. First let’s create the primary Job (the duty to detect the occasions). We have to instantiate a TaskUpdater object (a helper class for brokers to publish updates to a activity’s occasion queue), then submit the duty and announce we’re engaged on it with the start_work() methodology:

from a2a.server.agent_execution import AgentExecutor

class EventAgentExecutor(AgentExecutor):  
    async def execute(self, context: RequestContext, event_queue: EventQueue):  
        task_updater = TaskUpdater(event_queue, context.task_id, context.context_id)  
        task_updater.submit()  
        task_updater.start_work()

The message the consumer will ship to the agent will seem like this:

send_message_payload = {  
        'message': {  
            'function': 'consumer',  
            'elements': [{'type': 'text', 'text': f'it has an event!'}],  
            'messageId': uuid4().hex,  
        }  
    }

A Half represents a definite piece of content material inside a Message, representing exportable content material as both TextPart, FilePart, or DataPart. We’ll use a TextPart so we have to unwrap it within the executor:

from a2a.server.agent_execution import AgentExecutor

class EventAgentExecutor(AgentExecutor):  
    async def execute(self, context: RequestContext, event_queue: EventQueue):  
        task_updater = TaskUpdater(event_queue, context.task_id, context.context_id)  
        task_updater.submit()  
        task_updater.start_work()

        await asyncio.sleep(1) #let's faux we're truly doing one thing

        user_message = context.message.elements[0].root.textual content # unwraping the TextPart

Time to create the tremendous superior logic of our agent. If the message doesn’t have the string “occasion” we don’t have to name the Alert Agent and the duty is completed:

from a2a.server.agent_execution import AgentExecutor

class EventAgentExecutor(AgentExecutor):  
    async def execute(self, context: RequestContext, event_queue: EventQueue):  
        task_updater = TaskUpdater(event_queue, context.task_id, context.context_id)  
        task_updater.submit()  
        task_updater.start_work()

        await asyncio.sleep(1) #let's faux we're truly doing one thing

        user_message = context.message.elements[0].root.textual content # unwraping the TextPart

        if "occasion" not in user_message:  
            task_updater.update_status(  
                TaskState.accomplished,  
                message=task_updater.new_agent_message(elements=[TextPart(text=f"No event detected")]),
            )

Creating an A2A Consumer for the Consumer

Let’s create an A2A Consumer so we will check the agent as it’s. The shopper makes use of the get_client_from_agent_card_url() methodology from A2AClient class to (guess what) get the agent card. Then we wrap the message in a SendMessageRequest object and ship it to the agent utilizing the send_message() methodology of the shopper. Right here is the complete code:

import httpx  
import asyncio  
from a2a.shopper import A2AClient  
from a2a.sorts import SendMessageRequest, MessageSendParams  
from uuid import uuid4  
from pprint import pprint
  
async def major():    
    send_message_payload = {  
        'message': {  
            'function': 'consumer',  
            'elements': [{'type': 'text', 'text': f'nothing happening here'}],  
            'messageId': uuid4().hex,  
        }  
    }  

    async with httpx.AsyncClient() as httpx_client:  
        shopper = await A2AClient.get_client_from_agent_card_url(  
            httpx_client, 'http://localhost:10008'  
        )  
        request = SendMessageRequest(  
            params=MessageSendParams(**send_message_payload)  
        )  
        response = await shopper.send_message(request)  
        pprint(response.model_dump(mode='json', exclude_none=True))  
  
if __name__ == "__main__":  
    asyncio.run(major())

That is what occurs within the terminal that’s operating the EventAgent server:

And that is the message the shopper sees:

The duty to detect the occasion was created and no occasion was detected, good! However the entire level of A2A is to make Brokers talk with one another, so let’s make the Occasion Agent speak to the Alert Agent. 

Making the Occasion Agent speak to the Alert Agent

To make the Occasion Agent speak to the Alert Agent the Occasion Agent will act as a shopper as nicely:

from a2a.server.agent_execution import AgentExecutor

ALERT_AGENT_URL = "http://localhost:10009/" 

class EventAgentExecutor(AgentExecutor):  
    async def execute(self, context: RequestContext, event_queue: EventQueue):  
        task_updater = TaskUpdater(event_queue, context.task_id, context.context_id)  
        task_updater.submit()  
        task_updater.start_work()

        await asyncio.sleep(1) #let's faux we're truly doing one thing

        user_message = context.message.elements[0].root.textual content # unwraping the TextPart

        if "occasion" not in user_message:  
            task_updater.update_status(  
                TaskState.accomplished,  
                message=task_updater.new_agent_message(elements=[TextPart(text=f"No event detected")]),
            )
        else:
            alert_message = task_updater.new_agent_message(elements=[TextPart(text="Event detected!")])

            send_alert_payload = SendMessageRequest(  
                params=MessageSendParams(  
                    message=alert_message  
                )  
            )  

            async with httpx.AsyncClient() as shopper:  
                alert_agent = A2AClient(httpx_client=shopper, url=ALERT_AGENT_URL)  
                response = await alert_agent.send_message(send_alert_payload)  

                if hasattr(response.root, "outcome"):  
                    alert_task = response.root.outcome  
                    # Polling till the duty is completed
                    whereas alert_task.standing.state not in (  
                        TaskState.accomplished, TaskState.failed, TaskState.canceled, TaskState.rejected  
                    ):  
                        await asyncio.sleep(0.5)  
                        get_resp = await alert_agent.get_task(  
                            GetTaskRequest(params=TaskQueryParams(id=alert_task.id))  
                        )  
                        if isinstance(get_resp.root, GetTaskSuccessResponse):  
                            alert_task = get_resp.root.outcome  
                        else:  
                            break  
  
                    # Full the unique activity  
                    if alert_task.standing.state == TaskState.accomplished:  
                        task_updater.update_status(  
                            TaskState.accomplished,  
                            message=task_updater.new_agent_message(elements=[TextPart(text="Event detected and alert sent!")]),  
                        )  
                    else:  
                        task_updater.update_status(  
                            TaskState.failed,  
                            message=task_updater.new_agent_message(elements=[TextPart(text=f"Failed to send alert: {alert_task.status.state}")]),  
                        )  
                else:  
                    task_updater.update_status(  
                        TaskState.failed,  
                        message=task_updater.new_agent_message(elements=[TextPart(text=f"Failed to create alert task")]),  
                    )

We name the Alert Agent simply as we known as the Occasion Agent because the consumer, and when the Alert Agent activity is completed, we full the unique Occasion Agent activity. Let’s name the Occasion Agent once more however this time with an occasion:

The wonder right here is that we merely known as the Alert Agent and we don’t have to know something about the way it alerts the consumer. We simply ship a message to it and await it to complete.

The Alert Agent is tremendous much like the Occasion Agent. You may verify the entire code right here: https://github.com/dmesquita/multi-agent-communication-a2a-python

Last Ideas

Understanding the way to construct multi-agent programs with A2A is perhaps daunting at first, however in the long run you simply ship messages to let the brokers do their factor. All you want to do to combine you brokers with A2A is to create a category with the agent’s logic that inherit from the AgentExecutor and run the agent as a server.

I hope this text have helped you in your A2A journey, thanks for studying!

References

[1] Padgham, Lin, and Michael Winikoff. Growing clever agent programs: A sensible information. John Wiley & Sons, 2005.

[2] https://github.com/google/a2a-python

[3] https://github.com/google/a2a-python/tree/main/examples/google_adk

[4] https://developers.googleblog.com/en/agents-adk-agent-engine-a2a-enhancements-google-io/