Google just announced a new open protocol called Agent2Agent (A2A). This protocol allows different AI agents (or digital assistants) to talk to each. It ensures a secure and reliable communication, even if they come from different companies or platforms.

ServiceNow is one of the partners joining this initiative, along with 50+ others.

What is A2A protocol?

A2A is built on trusted, open standards. The protocol avoids shared memory or tooling. Instead, agents communicate by exchanging context, state, and capabilities — using a clean, async-first approach designed for flexibility and scale. When implemented, the AI agents can:

• Discover each other automatically
• Coordinate tasks together
• Communicate asynchronously (real-time or in the background)
• Authenticate securely (enterprise-grade security)
• Interact in many ways, not just text — e.g., audio, video, forms, or UI components

In simple words, you give digital workers a universal language so they can collaborate, no matter who made them.

A2A Specifications (are not yet fully clear)

​The Agent2Agent (A2A) Protocol facilitates secure and vendor-neutral communication between AI agents. According to the Google Developers Blog, the key specifications of A2A include multiple layers of capabilities. The specifications are still evolving but from the similar concepts it might include following in the future.

With a dynamic agent discovery, agents will locate and identify each other autonomously within a network. Means, there should be a central or decentralized registry that will act like a “phone book” where agents register themselves with metadata incl. name or ID, capabilities (e.g., “can summarize documents”), protocol versions supported, health status, availability, authentication/public key info, costs etc. Once discovered and authenticated, agents connect via the A2A protocol (usually HTTP/gRPC with async support) to exchange information, ask for information, trigger actions etc.

Open standards for connecting Agents (Source: Google)

The collaboration between AI agents can trigger and orchestrate complex workflows that typically include dividing responsibilities, subtasks, parallel and subsequent execution. Rather than relying on a single system to handle everything, specialized agents can coordinate and delegate work dynamically. This kind of workflow will even more improve efficiency, scalability, flexibility, and potentially lower costs. Similar to how teams operate in business: each agent / worker contributes its expertise, and the results are seamlessly integrated to deliver complete, end-to-end outcomes.

Communication and results can be of different types. In Gen AI you call it multimodal. This so-called modality-agnostic interaction allows AI agents to collaborate using a variety of formats—such as text, forms, audio, video, and user interface elements. The agents can adapt to different user needs, devices, and environments, delivering a more seamless and accessible experience. As a result, the A2A protocol enables richer, more intuitive workflows.

What is the difference between MCP and A2A?

Anthropic introduced the Model Context Protocol (MCP) in November 2024, which standardizes how AI applications (particularly those utilizing large language models) connect with external data sources and tools. It provides a unified interface that allows AI systems to access diverse data repositories, APIs, and services. MCP simplifies integrations and reduces the need for custom connectors. By using MCP, LLM-based applications can retrieve and utilize external context and improve their responses and functionalities.

Below I tried to put key differences together:

	Model Context Protocol	Agent2Agent
Scope of Integration	Connects AI models and applications to external data sources and tools. It enhances the model’s access to relevant context and functionalities.	Enables communication and collaboration between autonomous AI agents. It let these applications work together across different platforms and services.
Primary Functionality	Standardizes the interface for data retrieval and tool integration, so AI models can incorporate external information into their processing.	Builds a framework for agent interoperability, and how agents discover each other, share capabilities, and coordinate actions.
Use Cases	This covers scenarios where external data or services need to be accessed and utilised, such as retrieving real-time information or leveraging specialized tools.	A2A is suited for environments where multiple AI agents must collaborate. For example, orchestrating tasks across different enterprise applications or coordinating multi-agent workflows.

So, yes, they can enable similar things, but the idea, purpose, and scope behind them are fundamentally different.

MCP = Model-to-Tool/Data Integration. It’s like plugging apps into a smart assistant (like Siri or Alexa) so it can do more things.

A2A = Agent-to-Agent Communication. It’s more like giving every employee in a digital company a shared language and rules to collaborate — regardless of which vendor built them.

Or, in other words, you might use MCP inside an A2A-enabled agent as a way for that agent to fetch external data or trigger a tool. MCP can be “inside the agent,” while A2A governs how agents interact with each other.

Possible use cases for A2A communication

In a field service scenario, for example, AI agents collaborate to assist a technician through a multimodal experience using the A2A protocol. The following workflow demonstrates how multiple agents—each specializing in a different function—can coordinate in real-time using text, voice, video, and interactive UI components.

Imagine a field technician is on-site diagnosing an issue with industrial machinery. She/he speaks into her wearable device:

“Show me the latest maintenance log and initiate a repair request.”

An AI voice assistant agent (audio) captures the command and:

1. Sends the query to a data retrieval agent that returns the maintenance log in text format (e.g. LangChain)
2. Another agent triggers a form-based UI on her tablet to submit a repair request (e.g. ServiceNow Now Assist)
3. A video tutorial agent suggests a how-to clip embedded in the interface (e.g. Microsoft Power Virtual Agents + SharePoint)
4. She completes the form and sends a voice confirmation: “Submit and notify maintenance team. (e.g. ServiceNow Now Assist)

Each agent handles its piece using different modalities—audio, text, UI, video—while working together seamlessly.

This use case highlights the power of modality-agnostic, multi-agent orchestration. It will improve speed, accuracy, and user experience in the field, while maintaining security, scalability, and enterprise readiness. It also illustrates how leading platforms such as LangChain, ServiceNow, and Google Cloud can be combined to deliver intelligent, human-like support across the enterprise.

Are the enterprises ready for A2A protocol?

​It is not yet fully clear to me if the protocol will be able to meet the stringent requirements of enterprise environments. This includes robust security measures, compliance with industry standards, scalability to handle large workloads, and comprehensive support for governance and auditing. But also costs – who will pay and how to let AI agents talk to each other and ask for help.

However, the available information describes comprehensive concepts. Additionally, it was introduced by Google Cloud and supported by over 50 partners—including Atlassian, Box, Cohere, Intuit, LangChain, MongoDB, PayPal, Salesforce, SAP, ServiceNow, UKG, Deloitte, and Workday. Big names behind this initiative.

Open questions

There are some questions that are very relevant in the enterprise context.

• Will agents have billing APIs?
• How do you monitor agent-to-agent abuse or loops?
• Will cost models evolve like cloud billing (pay-per-call, compute-time, etc.)?
• How IT, CISO, Procurements and others will manage the governance around AI agents that communicate via A2A protocol?

Do you know the answers? What do you think about the above questions?

Summary

The Agent2Agent (A2A) protocol is an exciting step towards creating an open, interoperable AI ecosystem across enterprise platforms. It allows secure, asynchronous communication between AI agents, no matter who made them, and supports dynamic task orchestration in different ways (like text, voice, UI, video) and modular system design.

The opportunities for scaling up intelligent automation across different departments are huge. You just need to let every department develop its specific AI agents and let them talk.

A2A is like a modern-day game-changer, just like how cloud computing, microservices, Robotic Process Automation (RPA) and API-first architectures transformed everything we knew. Just as cloud adoption decoupled infrastructure from physical hardware and microservices broke down monolithic systems into scalable, independently deployable services, A2A is doing the same for intelligent automation.

It’s an exciting time: agents can now work modularly, communicate seamlessly, and scale across the enterprise stack.

There are still some risks, but these can be mitigated. The ecosystem is still in its early stages, so there aren’t many real-world examples yet, and the standards are still developing, but this is exciting because it means the potential is huge. Integration into existing enterprise environments may require some non-trivial architecture adjustments, but these are easily overcome. While security is a top priority, we’re confident that governance, compliance, and operational readiness are up to scratch.

So, let’s proceed with a sense of cautious optimism! Starting a small pilot in a non-critical area is a great way to test out the integration, vendor support, and security, before rolling it out more widely.

Kostya Bazanov, Managing Director, Apr 10, 2025

Eager to take the next step? Contact us today!

Privacy policy (GDPR)

* Required fields

Submit