Agent Protocols

Author(s): Burak Degirmencioglu

Originally published on Towards AI.

The rapid advancements in the field of artificial intelligence are moving beyond intelligent systems that work in isolation, opening the door to complex ecosystems of AI agents that can communicate and collaborate with one another. For these ecosystems to reach their full potential, a common language — agent protocols — must be understood and adopted.

In this article, I will explore what these protocols are, why they have become so important, and the main protocols shaping the ecosystem.

Why is Communication in the Agent Ecosystem So Fragmented?

Although today’s AI models have incredible capabilities on their own, they often operate in isolation. Each model has its own unique data formats, command sets, and operational logic. This prevents AI agents developed on different platforms and serving different purposes from communicating with each other, exchanging information, and achieving common goals. This complex situation is like bringing together incompatible operating systems or people who speak different languages. This leads to losses in efficiency and serious difficulties in system integration.

From Communication Fragmentation to Architectural Clarity: Agent Protocols

To solve this chaos, a common “language” for AI agents became essential, much like TCP/IP enabled communication between different computers on the internet. The rise of agent protocols is based on this very need. These protocols enable AI agents created by different developers, companies, or platforms to communicate, share their tasks, and share their capabilities within the same format and rules. This paves the way for creating much more complex and powerful systems by combining the capabilities of individual agents.

For example, a text generation agent can seamlessly interact with an image processing agent to create a comprehensive report containing both text and visuals, all based on the user’s request.

After understanding the importance of this coordination between agents, let’s take a closer look at one of the leading protocols in this field: the Agent-to-Agent (A2A) protocol.

What is the A2A (Agent-to-Agent) Protocol and How Does It Work?

As the name suggests, the A2A protocol was developed to enable AI agents to communicate with each other directly and meaningfully. Developed by Google and adopting an open-source approach, this protocol’s main goal is interoperability among agents. The protocol defines a standard interface and message format for agents to send data, delegate tasks, and collaborate. This allows agents to introduce their capabilities to each other, saying something like, “Hello, I’m a weather forecasting agent and I have this information.”

In this way, a travel planning agent can communicate seamlessly with a hotel agent for reservations, a taxi agent for transportation, and another agent for weather information.

The A2A protocol sets specific rules for the content of messages, the identity of the sending and receiving agents, and the purpose of the message. This allows agents to easily discover each other and query their capabilities. Its core components include message envelopes, content formats, and routing mechanisms. This establishes a reliable and consistent communication channel between agents.

In addition to A2A, another important protocol designed for decentralized networks is the ANP (Agent Network Protocol).

How Do Agents Interact in Decentralized Networks: What is ANP?

The Agent Network Protocol (ANP) is a protocol designed to provide secure and efficient communication, especially in decentralized, distributed, and peer-to-peer agent networks. Unlike A2A, it aims to allow agents to connect directly with each other without needing a central structure. ANP ensures the network has a layered architecture. These layers include Discovery and Identity mechanisms that manage agent identities, provide security, and allow agents to find other agents on the network.

These mechanisms allow agents to dynamically join and leave the network and guarantee that each agent has a unique identity. For example, multiple e-commerce agents operating on an ANP network can automatically communicate with each other to find the product a user is looking for at the best price. Each agent can query other agents on the network to collect data such as stock, price, and delivery information and collaborate to present the best offer.

Another important protocol that manages this interaction between agents is the Model Context Protocol (MCP).

Why was the Model Context Protocol (MCP) Developed and How Does It Work?

The Model Context Protocol (MCP) is a protocol that focuses on the need for context and state to be preserved and transmitted between AI models, especially large language models (LLMs). Beyond inter-agent communication, this protocol makes it possible for a model to transfer the context of a specific task or conversation flow to another model or agent. This allows agents to continue a process from where they left off, preserving all previous information and context instead of starting from scratch every time they hand off a task to a different model.

The core components of MCP include Agent State and Tool/Function Schema definitions. Agent State is a data structure that holds an agent’s current context, variables, and progress. The Tool/Function Schema defines the tools or functions that an agent can use.

For example, a customer service agent can share the context of a conversation with an expert system agent to solve a complex problem, and the expert system can use this context to continue the process from where it left off.

What is the Purpose of the A2H (Agent-to-Human) Protocol in Agent-Human Interaction?

AI agents should not only communicate among themselves but also effectively with humans. The Agent-to-Human (A2H) protocol focuses on this specific type of interaction. This protocol creates a standard framework for an AI agent to correctly understand a human user’s commands, intent, and feedback, and respond to the human in an understandable way. The A2H protocol defines how agents interact with user interfaces (UI) and user experience (UX).

The main components of this protocol include message formats and UI schemas. Message formats determine how an agent processes text, voice, or visual commands received from a human, while UI schemas standardize which buttons, text fields, or visuals the agent can display on a website or mobile application. This allows an AI agent to present a user’s request in the most appropriate and understandable way.

In addition to these protocols that facilitate communication among agents and with humans, the Agent Communication Protocol (ACP), a general-purpose communication protocol, also holds an important place.

What Problems Does the Agent Communication Protocol (ACP) Solve?

The Agent Communication Protocol (ACP) is a general-purpose protocol that regulates information exchange and interaction among AI agents. ACP is primarily designed to enable agents to complete complex tasks collaboratively. This protocol allows agents to share tasks, update their statuses, and come together to form larger systems. ACP provides a universal communication standard that is not specific to a particular platform or agent. This allows agents created by different developers to come together and work in a common ecosystem.

For example, a logistics agent can communicate with an inventory management agent to query stock status, then communicate with a shipping agent to create a shipment, and finally, talk to a finance agent to complete the payment process. Each of these processes follows a seamless flow thanks to ACP.

While each of these different protocols has its own purpose and strengths, considering and comparing them together gives us a clearer vision of the future of the AI ecosystem.

What is the Relationship Between Different Agent Protocols and How Do They Integrate?

Each protocol focuses on a different need within the AI ecosystem. A2A and ANP generally focus on direct and peer-to-peer communication between agents, while MCP focuses on preserving and transferring the context of a conversation or task. A2H standardizes human-agent interaction, and ACP provides a more general communication framework.

For example, an agent on an ANP network can use A2A to communicate with another agent to start a task, use MCP to preserve the context while performing this task, and use A2H to provide feedback to a human at the end of the task. These protocols work in a complementary manner, enabling the development of more comprehensive and flexible AI systems.

In short, the future of communication between AI agents depends on the integration of these different protocols. These protocols enable agents serving different purposes to come together and complete complex tasks that they could not do alone.

Do you think these protocols will merge into a single “universal” protocol in the future, or will each continue to develop within its own niche? Feel free to share your thoughts in the comments and stay tuned for developments on this topic.

Links

IBM – AI Agent Protocols

ANP — Github Repo

MCP

Google — A2A

Everest Group — Agent protocols

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI