Choosing the Right LLM AI Agent Protocol: A Four‑Category Guide

Overview of LLM AI Agent Communication Protocols

The article systematically reviews existing LLM AI Agent communication protocols and groups them into four major categories to help users and developers select the most suitable protocol for specific application scenarios.

1. AI Agent Protocol Development

Definition of AI Agent Protocol

An AI Agent protocol is a standardized framework that defines the rules, formats, and procedures for structured communication between agents and between agents and external systems. Compared with traditional interaction mechanisms such as APIs, GUIs, or XML, protocols offer significant advantages in efficiency, scope, standardization, and AI‑native capabilities.

Key Functions of Agent Protocols

Interoperability : Enables seamless collaboration among heterogeneous agents with different architectures.

Standardized Interaction : Allows agents to integrate and extend functionality easily by incorporating new tools, APIs, or services.

Security and Governance : Provides built‑in mechanisms to manage agent behavior within clearly defined security parameters.

Reduced Development Complexity : Abstracts interaction logic, letting developers focus on core agent capabilities.

Facilitating Collective Intelligence : Shares insights and coordinates actions through standardized communication channels, achieving results unattainable by a single architecture.

2. AI Agent Protocol Classification Framework

The article proposes a two‑dimensional classification framework that divides protocols into Context‑Oriented Protocols and Inter‑Agent Protocols, each further split into General‑Purpose and Domain‑Specific types.

1. Context‑Oriented Protocols

General‑Purpose Protocols

MCP (Model Context Protocol) : Proposed by Anthropic, a universal context‑acquisition protocol that lets AI agents interact with external resources (data, tools, services) via a client‑server architecture, decoupling tool calls from LLM responses and enhancing data security and privacy.

Domain‑Specific Protocols

agents.json : Introduced by WildCardAI, an OpenAPI‑based machine‑readable contract format that bridges traditional APIs with AI agents, supporting AI‑compatible interface declarations, authentication schemes, and multi‑step workflows.

2. Inter‑Agent Protocols

General‑Purpose Protocols

ANP (Agent Network Protocol) : Community‑driven protocol for interoperability among diverse agents, building an open, secure, and efficient collaboration network with cross‑domain communication and decentralized identity authentication.

A2A (Agent2Agent Protocol) : Developed by Google for complex problem solving and collaboration among enterprise‑internal agents, supporting asynchronous workflows and multimodal interaction.

AITP (Agent Interaction & Transaction Protocol) : Proposed by the NEAR Foundation, enables secure communication, negotiation, and value exchange between agents, especially across trust boundaries.

AConP (Agent Connect Protocol) : Cisco’s standard interface for invoking and configuring agents.

AComP (Agent Communication Protocol) : From Al and Data, standardizes communication between agents to promote automation and collaboration.

Agora : Oxford University’s meta‑protocol that allows agents to select different communication protocols based on context.

Domain‑Specific Protocols

LMOS (Language Model Operating System) : Eclipse Foundation’s ecosystem for discovering, interacting with, and interoperating internet agents.

Agent Protocol : AI Engineer Foundation’s standard for console‑to‑AI‑agent communication.

LOKA : CMU’s decentralized framework for trust and ethical coordination among knowledge agents.

PXP (Predict and eXplain Protocol) : BITS Pilani’s protocol focusing on bidirectional explainability in human‑machine interaction.

CrowdES : GIST.KR’s protocol for robot agents to simulate real‑world crowd dynamics.

SPPs (Spatial Population Protocols) : University of Liverpool’s solution for distributed localization among anonymous robots.

3. Case Study: Five‑Day Beijing‑to‑New York Travel Planning

The article compares four AI agent protocols—MCP, A2A, ANP, and Agora—using a travel‑planning scenario.

1. MCP: Single Agent Calls All Tools

Suitable for well‑defined tasks with stable external service interfaces; not ideal for dynamic environments.

User sends a travel request to the MCP Travel Client.

MCP Travel Client directly calls Flight Server, Hotel Server, and Weather Server to obtain respective information.

After receiving responses, the client aggregates them into a complete travel plan.

2. A2A: Complex Collaboration Within an Enterprise

Fits scenarios where multiple specialized agents communicate directly inside an organization, relying on stable internal infrastructure.

User submits a travel request to the A2A Travel Planner.

The planner distributes the task to specialized agents (Flight, Hotel, Weather).

Agents communicate directly (e.g., Flight Agent queries Weather Agent for weather data).

Agents return results to the planner, which aggregates the final itinerary.

3. ANP: Cross‑Domain Agent Protocol

Designed for inter‑organization collaboration requiring clear protocols and security mechanisms.

User requests a travel plan from the ANP Travel Planner.

The planner assigns tasks to agents from different organizations.

Agents interact across organizational boundaries (e.g., Flight Agent with Weather Agent).

Results are returned to the planner for final aggregation.

4. Agora: Natural‑Language‑to‑Protocol Generation

Focuses on user interaction, converting natural language into structured protocols for efficient user‑centric scenarios.

User issues a natural‑language request such as “Plan a five‑day trip from Beijing to New York”.

Agora’s NLU module parses the request and extracts key information (origin, destination, duration, budget).

The protocol generation module translates this information into standardized protocols (Flight, Hotel, Weather).

The protocol distribution module forwards each protocol to the corresponding specialized agents.

Agents respond according to the protocols, and results are compiled.

https://arxiv.org/pdf/2504.16736</code><code>Survey of AI Agent Protocols

Source: PaperAgent public account.