A Survey of AI Agent Protocols

📝 Paper Summary

Agent Communication Protocols Multi-Agent Systems

This survey provides a comprehensive analysis and two-dimensional classification of AI agent protocols to address the fragmentation of current agent ecosystems and enable standardized, scalable collaboration.

Core Problem

The rapid deployment of LLM agents has led to a fragmented ecosystem where agents cannot easily communicate with external tools, data sources, or other agents due to a lack of standardized protocols.

Why it matters:

Current isolation limits agents' ability to tackle complex, real-world tasks that require collaboration or diverse tool usage
Lack of interoperability prevents the emergence of 'collective intelligence' where specialized agents and intelligent tools combine capabilities
Similar to the pre-Internet era, the absence of unified standards (like TCP/IP) hinders the scalability and global connectivity of intelligent systems

Concrete Example: Currently, an agent from one vendor cannot easily collaborate with an agent from another provider because they lack a shared 'grammar' for exchange. This mirrors the early Internet's incompatible systems before TCP/IP; without a protocol, a 'customer service' agent cannot seamlessly hand off a complex query to a specialized 'data analysis' agent from a different platform.

Key Novelty

Systematic Taxonomy of Agent Protocols

Proposes a two-dimensional classification framework: distinguishing Context-oriented vs. Inter-agent protocols, and General-purpose vs. Domain-specific protocols
Conceptualizes protocols not just as APIs, but as a foundational grammar that enables 'collective intelligence' and separates interaction logic from agent core functionality

Breakthrough Assessment

7/10

While a survey rather than a new algorithm, it addresses a critical infrastructure gap (standardization) necessary for the next leap in agentic AI, providing a much-needed organizational framework.

⚙️ Technical Details

Problem Definition

Setting: Defining standardized frameworks (protocols) for structured communication within the LLM agent ecosystem

Inputs: Heterogeneous LLM agents, external tools, and data sources

Outputs: Standardized, interoperable, and scalable communication channels

Comparison to Prior Work

vs. APIs: Protocols provide a standardized 'grammar' allowing dynamic integration of new tools without custom glue code for every integration
vs. XML/HTML: Protocols are designed for high-efficiency, complex AI-to-AI interaction rather than just rendering or basic data tagging

Limitations

The survey relies on the current nascent state of protocols, which are rapidly evolving
Quantitative performance comparisons (mentioned in abstract) are dependent on the specific implementations of the protocols which may vary by vendor
Adoption of a unified standard faces challenges from competing proprietary ecosystems (e.g., model-centric vs. open network approaches)

Reproducibility

Code: https://github.com/zoe-yyx/Awesome-AIAgent-Protocol

The paper is a survey. The authors provide a GitHub repository (https://github.com/zoe-yyx/Awesome-AIAgent-Protocol) which serves as a maintained list of protocols and resources discussed in the paper.

📊 Experiments & Results

Evaluation Setup

Qualitative comparison of interaction mechanisms and survey of existing protocols

Metrics:

Efficiency
Scalability
Standardization
Operational Scope
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

Protocols offer a distinct advantage over APIs and GUIs by combining high efficiency, extensive operational scope, and native AI compatibility
The landscape is currently split between model-centric (MCP), enterprise-focused (A2A), and open network (ANP) strategic paradigms
Standardization is identified as the prerequisite for 'collective intelligence', where tool intelligence and agent intelligence can merge dynamically
Future protocols must support adaptability, privacy preservation, and group-based interaction to handle complex agent ecosystems

📚 Prerequisite Knowledge

Prerequisites

Understanding of LLM Agent architecture (Memory, Planning, Tools)
Familiarity with API vs. Protocol distinctions
Basic knowledge of distributed systems

Key Terms

LLM Agent: An AI system integrating a Large Language Model with memory, planning, and tool-use capabilities to perform autonomous actions

MCP: Model Context Protocol—Anthropic's standard for context acquisition between agents and external resources

ANP: Agent Network Protocol—A decentralized protocol aiming to establish an open agent internet for widespread interoperability

A2A: Agent-to-Agent—Google's enterprise-focused protocol prioritizing integration, security, and governance

Context-oriented Protocol: Protocols designed to standardize how agents acquire and manage context from external resources (e.g., data, tools)

Inter-agent Protocol: Protocols designed to facilitate collaboration and message exchange between different agents

Collective Intelligence: Emergent intelligence arising from the collaboration of multiple specialized agents and tools, greater than the sum of individual components