Google launches A2A protocol, ushering in the era of cross-platform collaboration among intelligent agents

2025 is the first year of intelligent agents. The explosion of intelligent agent applications based on large models has also brought more interoperability requirements. After Athropic launched the MCP protocol to allow intelligent agents to better integrate with external resources, tools, and API interfaces, Google launched the open A2A ( Agent2Agent ) protocol in April , aiming to improve the interoperability between heterogeneous intelligent agents across platforms. Originally, a single intelligent agent fought alone. If a user used multiple intelligent agents to complete a task, it was necessary to manually connect different intelligent agents (such as copying and pasting, exporting and importing, etc.). Now with the A2A protocol, intelligent agents can collaborate with each other to complete more complex tasks together. As an open protocol, the A2A protocol allows intelligent agents from different frameworks and suppliers to collaborate with each other, solving the challenges faced by enterprises when adopting multi-agent systems.

Although it has just been launched, more than 50 industry-leading companies have announced their support for the A2A protocol . These include two types of companies: one is technology partners , such as Atlassian , Box , Cohere , Intuit , Langchain , MongoDB , PayPal , Salesforce , SAP , ServiceNow , UKG , Workday , etc., and the other is service providers , including Accenture , Boston Consulting Group (BCG) , Capgemini , Deloitte , Infosys , KPMG , McKinsey , etc. This approach is similar to Google's joint efforts with more than 80 companies to open source the Android system, hoping to create a more open intelligent ecosystem .

01 Core Concepts of A2A Protocol

The core concepts of the A2A  protocol include:

• Agent Card : A public metadata file that describes the agent's capabilities, skills, endpoint  URLs  , and authentication requirements. Clients use it for agent discovery.
  

• A2A  Server : An agent that implements the  A2A  protocol methods, receives requests and manages task execution.
  

• A2A  client :An application or other agent that uses  A2A  services and sends requests to  the A2A  server.
  

• Task : The basic unit of work . The client starts a task by sending a message.
  

• Message : represents a communication round between the client and the agent, and consists of multiple parts .
  

• Part : The basic unit of content in a message or artifact, which can be text, a file, or structured data.
  

• Artifact : The output generated by the agent during the task.

  
  

A task is a job that a user needs to complete using an agent system. The output of a task is an artifact , such as a research report that a user needs an agent to complete.

AI applications can use the A2A client to send a request and start a task. After receiving the request, the A2A server processes the task and can interact with multiple other agents to complete the task together. Each agent has an AgentCard to describe its identity, capabilities, and usage. The A2A server discovers agents through AgentCards and understands what each agent can do and how to use it.

Communication between clients and agents is done using messages. Messages are entities that contain any non-artifact content. This content can include agent thinking, user context, instructions, error messages, status updates, or metadata. Messages can contain multiple parts , each of which can be data in different formats, including text, files, structured data, etc.

02 Typical workflow of A2A protocol

The typical workflow of the A2A  protocol includes the following stages:

1. Discovery: The client obtains  the Agent Card from a known  URL  on the server .

   
   

2. Initialization: The client sends a request containing an initial user message and a unique task  ID  .

   
   

3. Processing: Streaming: The server sends task progress and artifact updates via  SSE  events. Non-streaming: The server processes the task asynchronously and returns the final task object.

   
   

4. Interaction: If the task requires input, the client sends subsequent messages with the same task  ID  .

   
   

5. Completion: The task eventually reaches a completed, failed, or cancelled state.

03 Communication mechanism and asynchronous support of A2A protocol

A2A  supports three communication mechanisms:

1. A2A supports a secure push notification mechanism, allowing the agent to send updates to the client without being connected;

2. The client and server can use the standard request / response model;

3. Clients and servers can also stream updates via SSE (Server Sent Events).

When pushing notifications, the agent needs to verify the identity of the notification service and authenticate with trusted credentials to ensure the security of the notification. The A2A  protocol has built-in support for authentication, authorization, security, and privacy. Each agent can declare its own authentication requirements in  the Agent Card , such as  OAuth2 , API Key  , or  JWT .

Based on the above communication mechanism, A2A  supports the client to poll when processing long-running tasks, and the agent can also push status updates to the client through  SSE  , so that it can support asynchronous methods . Because in large model applications, output takes a certain amount of time, if asynchronous methods are used, it can bring users a better experience.

04 Comparison between A2A protocol and MCP protocol

Both are communication protocols for intelligent agents. The A2A protocol focuses on communication between intelligent agents , while the MCP focuses on communication between intelligent agents and external resources, tools, and API interfaces . Through MCP , intelligent agents can call external tools or integrate with existing information systems.

Here is a comparison between A2A and MCP :

05 Multi-agent collaboration example

Consider a scenario of multi-agent collaboration. Before a user goes on a trip, he needs to make an itinerary, which requires the use of three agents: the itinerary assistant agent, the air / train ticket booking agent, and the hotel booking agent. Currently, the three agents are independent of each other. The user needs to use the itinerary assistant agent to first make a preliminary itinerary, then book air tickets and high-speed rail tickets through the air / train ticket booking agent, book a hotel through the hotel booking agent, and then update the air ticket, high-speed rail ticket, and hotel information to the itinerary.

If the A2A protocol is used , the user can directly make a task request to the itinerary assistant agent: Please help me plan a seven-day itinerary to Nanjing, Shanghai, and Hangzhou, and book air tickets / high-speed rail tickets and hotels. Then the three agents can collaborate to complete the above tasks. After the itinerary assistant agent completes the itinerary planning, it will send a booking request to the air ticket / train ticket booking agent and the hotel booking agent, and then wait for the processing results and update them to the itinerary, and finally feedback the processing results of the itinerary to the user. During the process, the itinerary assistant agent can also interact with the user at any time to obtain more information or confirm various processing options. The details are shown in the figure below.

Conclusion

The emergence of the A2A protocol facilitates collaboration between multiple agents across platforms, which will greatly promote the widespread application of agents. The role of A2A collaboration is like the interaction between people through language, which has become the foundation of human society. Different agents can communicate through the A2A protocol to build a distributed agent network , and greatly improve production efficiency through mutual collaboration .