The previous article just gave a preliminary introduction to the MCP protocol (Model Context Protocol) . When I was still learning MCP , I found that Google had launched the A2A (Agent2Agent) technology .

Today we are keeping up with the times and studying A2A

1. Introduction:

A2A is an open protocol developed under the leadership of Google . It aims to achieve standardized communication and collaboration between AI Agents from different manufacturers and frameworks , and solve the "information island" problem in the intelligent agent ecosystem.

The previous article compared MCP to "the Qin Shi Huang's unified wheel gauge and unified writing system" , and A2A can be seen as "the whole world uses the same world language and the same measurement system, etc. " - a real upgraded version. To put it simply, MCP enhances the capabilities of individual agents, while A2A allows agents to communicate and collaborate effectively, enhancing the capabilities of the "population". MCP is like humans learning to use tools to increase individual capabilities, while A2A is like humans having a unified language, increasing the capabilities of the "population"/

Although Google has made it clear that A2A is a supplement to Anthropic's MCP agreement, not a replacement , I personally believe , based on my past research on business history , that the two will eventually come together and develop into a whole that cooperates with each other, rather than always cooperating independently without crossing the line. I just don't know who will successfully integrate with whom in the end . However, there are still big differences between the two, and they are truly complementary.

The differences between the two are as follows:

• MCP : Solve the problem of how a single agent calls tools and APIs . The purpose is to enhance the agent's own capabilities by calling external tools and obtaining external data, and to practice its own professional capabilities.

• A2A : Solving the problem of how multiple agents can collaborate is to use other agents to help you "work" and enhance yourself through other agents. It is "I hire experts if I don't know how to do it" . It emphasizes task coordination between intelligent agents.

In other words, A2A is at a higher level of abstraction and is responsible for collaboration between agents, while MCP focuses on the connection between underlying tools and resources . For example, a task may be coordinated by A2A across multiple agents, and these agents call external tools through MCP to complete subtasks.

At present, A2A and MCP are complementary rather than competitive. A2A and MCP can work together to build a complex multi-agent system. In the same series of tasks, the two tasks are divided as follows: A2A coordinates the division of labor among multiple agents, and MCP provides tool support for each agent . For example, the sales agent calls CRM data through MCP, and then collaborates with the logistics agent through A2A to process orders.

Google's Agent Development Kit (ADK) supports both A2A and MCP. Developers can build multimodal collaborative agents that can collaborate across platforms and seamlessly access tools. This makes current collaboration possible and provides space for future integration .

At present, the emergence of the two is jointly promoting the openness and interoperability of the intelligent body ecosystem. It also marks the transition of the AI ​​industry from "fighting alone" to "combined army operations", accelerating the process of enterprise intelligence , which is definitely a leap forward.

2. Some original intentions and interpretations of A2A

After comparing the differences between the two (I was confused about the differences for a long time, and when I asked the technical experts for advice, they wanted to hit me with their family secrets), let's focus on A2A. First, I will excerpt some of the key points that Google said when it released the product and my views on these key points:

1. “AI agents offer a unique opportunity to help people be more productive by autonomously handling many daily, repetitive or complex tasks. Today, businesses are increasingly building and deploying autonomous agents to help scale, automate, and optimize entire workflows—from ordering a new laptop, to assisting a customer service representative, to aiding supply chain planning.”

From this, we can easily see the judgment of large companies on future trends:

More and more industries and enterprises will deploy their own industry-specific and professional Agents to help the industry optimize the entire workflow, improve efficiency, and reduce costs. As I said in my previous article, improving productivity has always been the general trend of human development.

Google is probably more technologically savvy than us, which is why it launched something like this. So everyone should definitely pay attention to the opportunities of industry agents. This is very difficult, so the moat is very high. People with ability and determination can really start now.

2. "Today, we are launching a new open protocol, Agent2Agent (A2A). The protocol has received support and contributions from more than 50 technology partners, including  Atlassian ,  Box ,  Cohere ,  Intuit ,  Langchain ,  MongoDB ,  PayPal , Salesforce,  SAP ,  ServiceNow , UKG and Workday; as well as leading service providers, including Accenture, BCG, Capgemini, Cognizant, Deloitte, HCLTech, Infosys, KPMG, McKinsey, PwC, TCS, and Wipro. The A2A protocol will allow AI agents to communicate with each other, exchange information securely, and coordinate actions on a variety of enterprise platforms or applications. We believe that the A2A framework will bring great value to customers, and their AI agents will be able to work across their entire enterprise application environment."

What is the core here?

I think it is the data security endorsed by Google . This is very important. Because when each agent provided data before, they did not get any visible gains, and faced the risk of possible data leakage, so it was difficult to promote their collaboration and they could only fight on their own. But Google's endorsement is different, there is "traffic", and it also tells you that it is safe .

Just like what I said in my previous article, Tencent provides you with traffic and security, so you can make mini programs. Do it? Just do it with your eyes closed. Isn't Google's promise the same as Alibaba's launch of Alipay to ensure transaction security? So this milestone product is really not just a casual remark.

3. Implementation of A2A

Let's take a look at A2A's technical innovations. The key point is the component settings. Let's introduce them one by one:

Component 1: Agent Card

In the A2A specification, each Agent must publish an "Agent Card", which is equivalent to a " self and self-capability declaration ", describing its own capabilities, skills, endpoint URL and authentication requirements, so that other Agents can understand, select and use it.

The client uses it to discover the proxy service (of course, discovering and selecting the proxy service is the biggest technical difficulty at present, especially the selection of the proxy service is the most cutting-edge technology being tackled, so it is inconvenient to disclose the details) .

Component 2: Agent Server (A2A Server)

A proxy that exposes an HTTP endpoint and implements the A2A protocol methods, which receives requests and manages task execution.

Component 3: A2A Client

Used to send requests, such as:

TaskSend , TasksSendSubscribe , TaskStatusUpdateEvent (task status update) or TaskArtifactUpdateEvent (task artifact update).

Component 4: Task

One of the core concepts in A2A is Task . When Agent1 wants to complete something through another Agent2, Agent1 will send a "contract invitation application" (TaskSend) to Agent2. After Agent2 agrees and returns the agreement (Task status changes), the two parties establish a link and create a Task ID to track project progress, exchange data, and update task progress in real time until the Task is completed.

Component 5: Streaming and Push Notifications

If a Task is a long-term project, or the Task is complex and requires many rounds of communication between Agents, or if the remote Agent takes a long time to execute, the progress can be updated regularly to the initiator through the push notification mechanism, "telling the initiator not to worry, I'm still working on it."

For long-running tasks, a server that supports the Streaming function can use tasks/sendSubscribe to send information to the initiator. After the initiator receives the SSE (Server-Sent Events) information, it maintains a long-term connection, which includes TaskStatusUpdateEvent or TaskArtifactUpdateEvent messages, to provide real-time progress updates.

This can greatly improve the ability of asynchronous collaboration , so as to avoid the initiator Agent waiting in vain, or thinking that the remote Agent has not executed and looking for other remote Agents, causing unnecessary consumption.

Component 6: Artifacts

Artifacts are used to show the final results. It may be a generated report, a picture or other forms.

Component 7: Message and Part

Part  is the basic content unit in a message or artifact, which can be TextPart (text), FilePart (file with inline bytes or URI), DataPart (structured JSON data, such as a form), etc.

Message is used for various rounds of communication during the task process. It contains additional explanations, corrections, further requirements, etc. of the task details.

A Message contains multiple Parts.

4. The typical workflow of A2A is as follows:

1 Discovery: The client learns about the capabilities of other agents through their Agent Cards.

2 Start: The client sends a task request:

• Use tasks/send to process immediate tasks and return the final Task object.

• Use tasks/sendSubscribe to handle long-term tasks, and the server sends updates via SSE events

  
  

3 Processing: Server processing tasks

• Instant tasks, directly return results

• Long-term tasks, streaming updates

  
  

4 Interaction (optional, used when needed): If the task status is input-required, the client can send more messages to provide input using the same Task ID

5 Completed: The task reaches a terminal state, which can be completed, failed, or canceled.

ps: This process supports both simple tasks and complex tasks that require multiple interactions, and is very suitable for multimodal communication environments.

5.Supplement:

You have the following options for how to use A2A:

Google's  A2A protocol (Agent2Agent Protocol)  is mainly aimed at developers and enterprise-level application scenarios. Its core interaction logic is implemented through the standard interface defined by the protocol, and ordinary users do not need to download the client separately . If you need to interact with the A2A agent on your computer, you need to choose the following tools according to your specific role:

1. Developer Tools

Google ADK (Agent Development Kit) is preferred

• Supported languages : Python, Java, Go

2. End-user interaction method

1. Browser access

1. Scenario : Interact with the agent through a web application that integrates the A2A protocol

2. Desktop Client

• Download : Microsoft Developer Center

• Function : Supports local client of A2A protocol and provides visual dialogue interface

1. Microsoft Bot Framework

• Command Line Tool (CLI)

1. Applicable scenarios : Developer debugging or automated script calling. Ordinary users can consider the previous two methods.

2. Installation command : pip install google-a2a-cli

3. Enterprise-level deployment solution

1. A2A Server

1. Function : As a hub for intelligent communication, managing task distribution and state synchronization

2. Deployment requirements : Enterprise server or cloud host is required (recommended configuration of 4 cores and 8GB or more)

2. Security Tools

1. OAuth 2.1 client : for intelligent identity authentication (such as Keycloak, Auth0)

2. National cryptographic algorithm support : You need to download the national cryptographic SSL library (such as gmssl)

   
   

VI. Outlook




When open protocols such as A2A and MCP gradually unify communication standards, and everyone contributes, the emergence of a new AI Agent ecosystem is almost inevitable .




Each agent constantly enriches its knowledge and strengthens its own capabilities, which may be industry capabilities or enterprise capabilities. In short, it must be capabilities in vertical fields. Then the overall capabilities of the agent can be improved by several orders of magnitude. And this process is the best opportunity for ordinary people like us to enter the game. No matter what industry it is, as long as you are professional in the industry, you can AI the industry and become an industry agent. Just like making a website for all industries or enterprises, and making apps for all industries or enterprises, it felt so difficult at the time, but is it true that if you work hard, you will have a chance? We don’t have the opportunity to make large models, but the accumulation of industry knowledge is the result of countless days and nights, and AIizing them is our great opportunity.

The era from only being able to see the effects to actually using Agents is getting closer and closer. As I said when I wrote Manus, everyone can have "Iron Man's Jarvis", and the era of AI empowering humans is coming.