In the fast-paced world of artificial intelligence, a discreet yet crucial revolution is underway: that of autonomous agents. These digital entities capable of reasoning, deciding, and acting independently are gaining importance in modern businesses. But until now, a major challenge remained to be solved: how to enable these agents to communicate effectively with each other, especially when they are designed by different providers? This is where the A2A (Agent-to-Agent) protocol comes into play, recently unveiled by Google and already supported by around fifty major tech players.
From MCP to A2A: a natural evolution
In a previous article, we discussed the MCP (Model Context Protocol), designed to allow an artificial intelligence to use external tools like an arm that complements the “head” represented by the LLM (Large Language Model). MCP thus enables an AI to take actions in the digital world – opening a file, sending a Slack message, querying a Notion database, etc.
But an essential ingredient was missing: direct communication between agents. In a rich ecosystem, where we envision a multitude of specialized agents – in data, marketing, logistics, legal… – it becomes essential to make them collaborate to accomplish complex tasks.
That’s why Google, with the support of many partners, is now proposing the A2A protocol.
What is A2A?
The Agent-to-Agent (A2A) protocol is an open standard aimed at enabling AI agents to communicate with each other smoothly and securely, regardless of their underlying technology or the provider that created them. It is a common language for agents, designed to promote interoperability, cooperation, and coordination in complex environments.
This protocol addresses a strategic necessity: companies do not want a closed system where each agent only works with tools from the same provider. They want to be able to assemble modular and interoperable agents, each being the best in its field, but capable of working together.
One protocol, multiple partners
Google did not embark on this adventure alone. Companies like Salesforce, Atlassian, Cohere, Langchain, MongoDB, SAP, ServiceNow, as well as major consulting firms like Accenture, BCG, Deloitte, KPMG, PwC, have all contributed to designing a standard capable of meeting field needs.
The main principles of the A2A protocol
The design of the A2A protocol is based on five fundamental principles:
1. Promote agent capabilities
The goal is to allow agents to collaborate without necessarily sharing memory, context, or tools. An agent can thus delegate a task to another, even if they do not operate on the same framework.
2. Rely on existing standards
A2A uses proven and widely adopted technologies: HTTP, JSON-RPC, SSE (Server-Sent Events)… This facilitates its adoption in current IT systems.
3. Security by design
Like modern APIs, security is integrated at the core of the protocol. A2A supports robust authentication and authorization, with schemes similar to those defined in OpenAPI.
4. Support for long tasks
A2A is not designed solely for instant tasks. It also manages long-running processes, with real-time status updates, notifications, and continuous progress tracking.
5. Modality agnostic
The protocol is not limited to text. It supports multimedia content such as audio or video, opening the door to rich interactions between agents.
How does A2A work in practice?
The protocol defines a typical interaction between two types of agents:
- Client agent: The one who formulates a task and seeks another agent to carry it out.
- Remote agent: The one who receives the task, executes it, and returns a result.
This operation relies on several key concepts:
✅ Capability discovery
Each agent can publish an Agent Card in JSON format, describing what it can do. This allows a client agent to quickly identify which other agent is best suited for a given task.
🛠️ Task management
Exchanges between agents revolve around a central object: the task. This can be simple or complex, instantaneous or long. Once the task is completed, the result is encapsulated in what is called an artifact.
🤝 Collaboration
Agents can send each other messages, instructions, artifacts, or context at any time. This communication is asynchronous and can include regular status updates.
🧩 User experience negotiation
Each message contains parts of content with an explicit type (text, image, iframe…). This allows agents to agree on the best way to present results, depending on the available display capabilities.
A Concrete Example: Launching a Multichannel Campaign
A tech company wants to launch a new communication campaign to promote its latest product. It has an ecosystem of specialized AI agents: some are experts in data analysis, others in writing, design, media buying, and campaign management.
Thanks to the A2A protocol, these agents can cooperate autonomously and securely to set up a successful campaign, without a human having to coordinate each step.
Steps of collaboration between agents
1. Defining objectives
- Strategic marketing agent: receives the instruction “Launch a campaign for the new product X targeting SMEs in France.”
- It formulates an A2A task and consults the other available agents via their Agent Cards.
2. Data analysis
- It delegates to a data agent the task of identifying the most receptive segments by cross-referencing:
- existing customer data,
- search trends,
- performance of past campaigns.
→ The data agent returns a artifact: a JSON report detailing the target personas and the most effective channels.
3. Content creation
- The marketing agent sends a task to an AI writing agent, with the brief and target personas.
- The agent generates several texts tailored for LinkedIn, email, and Google Ads.
- Simultaneously, an AI design agent creates visuals suitable for different formats (banners, carousels, images for newsletters).
→ These contents are returned as multimodal artifacts (text + images), with metadata on the formats.
4. Campaign deployment
- A media planner agent selects the optimal channels: Google Ads, emailing, LinkedIn, and sends subtasks to the agents responsible for each channel.
- The distribution agents connect to the tools via API (Meta Ads, Mailchimp, etc.) and launch the campaign.
→ Each sub-agent sends status updates and KPIs via A2A to ensure real-time traceability.
5. Monitoring and optimization
- The marketing agent monitors results via an interactive dashboard fed live.
- It can ask the writer to test a new hook or the designer to change a visual based on performance (automated A/B testing).
Result:
- The campaign is launched in just a few hours.
- The agents were able to:
- share data,
- create assets,
- coordinate actions,
- all without manual intervention between each step.
The same campaign could be automatically replicated in other countries:
- The marketing agent sends a “localization” task to a linguistic agent.
- This agent interacts with other agents to adapt messages, formats, and channels to the local market.
Why it matters?
The launch of the A2A protocol marks a key milestone in the evolution of AI agents. It is not just a new technical standard, but a engine for future distributed intelligent systems.
For businesses:
- Time savings: Workflows become smoother and faster.
- Interoperability: No more silos, agents can cooperate across departments and tools.
- Reduced costs: Less specific development for each integration.
For developers:
- Standardization: A single way to make agents communicate, regardless of the framework.
- Open ecosystem: Ability to combine agents from different providers.
What’s next for A2A?
Google and its partners are planning a stable version of the protocol for production adoption by the end of the year. The entire project is open source, with specifications already published, code examples, and a dedicated site to track the project’s evolution and contribute.
This hints at a future where companies can compose agent ecosystems as easily as assembling Lego bricks, each bringing its specialty, all capable of understanding each other.
Conclusion
The A2A protocol fills an essential gap in the architecture of agent systems. It is not just about making artificial intelligences “talk” to each other, but enabling them to truly collaborate to automate, optimize, and transform practices in businesses.
While the MCP protocol gave an arm to AI, A2A now gives it a voice, an ear, and a network. Together, they form the foundation of a new generation of distributed, collaborative, and interoperable AI.
We may be on the brink of an era where intelligent agents are no longer just simple assistants, but true digital teammates, capable of acting in network to transform our daily lives.
