PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of AI Agent Protocols

Yingxuan Yang, Huacan Chai, Yuanyi Song, Siyuan Qi, Muning Wen, Ning Li, Junwei Liao, Haoyi Hu, Jianghao Lin, Gaowei Chang, Weiwen Liu, Ying Wen, Yong Yu, Weinan Zhang

arXiv:2504.16736v3cs.AI

TL;DR

LLM agents increasingly interact with tools, data sources, and other agents, but the absence of standardized protocols limits interoperability and scalability. This survey classifies agent protocols systematically, compares their key properties, and identifies characteristics and architectures for future communication infrastructures.

  • Problem

    The lack of standardized protocols makes it difficult for agents from different providers or architectures to interoperate with external resources and collaborate at scale.

  • Method

    The paper conducts a comprehensive survey, proposes a two-dimensional classification, and evaluates protocols across key dimensions including efficiency, scalability, and security.

  • Results

    The survey organizes existing protocols, compares their trade-offs across key performance dimensions, and identifies emerging directions for evolvable, privacy-aware, and group-coordinated protocols.

  • Takeaways & Limitations

    The survey provides a practical reference for selecting and designing agent protocols while highlighting layered architectures and collective intelligence infrastructures as future directions.

Abstract

from arXiv · show

The rapid development of large language models (LLMs) has led to the widespread deployment of LLM agents across diverse industries, including customer service, content generation, data analysis, and even healthcare. However, as more LLM agents are deployed, a major issue has emerged: there is no standard way for these agents to communicate with external tools or data sources. This lack of standardized protocols makes it difficult for agents to work together or scale effectively, and it limits their ability to tackle complex, real-world tasks. A unified communication protocol for LLM agents could change this. It would allow agents and tools to interact more smoothly, encourage collaboration, and triggering the formation of collective intelligence. In this paper, we provide the first comprehensive analysis of existing agent protocols, proposing a systematic two-dimensional classification that differentiates context-oriented versus inter-agent protocols and general-purpose versus domain-specific protocols. Additionally, we conduct a comparative performance analysis of these protocols across key dimensions such as security, scalability, and latency. Finally, we explore the future landscape of agent protocols by identifying critical research directions and characteristics necessary for next-generation protocols. These characteristics include adaptability, privacy preservation, and group-based interaction, as well as trends toward layered architectures and collective intelligence infrastructures. We expect this work to serve as a practical reference for both researchers and engineers seeking to design, evaluate, or integrate robust communication infrastructures for intelligent agents.

A Survey of AI Agent Protocols

This section identifies the paper’s authors, contact affiliations, and key terms.

  • The paper lists Yingxuan Yang, Huacan Chai, Yuanyi Song, and additional coauthors.
  • The listed contact affiliation uses the sjtu.edu.cn domain.
  • The key terms are AI Agent Protocol, AI Agent, Agent Protocol Evaluation, and LLMs.

1 Introduction

LLM agents are increasingly used across industries, but heterogeneous systems lack standardized interaction protocols. This survey classifies existing protocols, compares their properties, and identifies directions for future agent communication infrastructures.

  • LLM agents are deployed in customer service, content creation, data analysis, and medical assistance, while interacting with external data sources, tools, and online agents.
  • The absence of standardized protocols creates incompatibility across vendors, tools, and agent architectures, limiting interoperability, network scalability, and complex real-world problem solving.
  • The paper provides a comprehensive analysis and comparative evaluation of agent protocols across dimensions including security, scalability, and latency.
  • The survey proposes a two-dimensional classification distinguishing context-oriented from inter-agent protocols and general-purpose from domain-specific protocols.
  • Future protocols are envisioned to support adaptability, privacy preservation, group-based interaction, layered architectures, and collective intelligence infrastructures.

2 Preliminaries

The preliminaries define LLM agents as autonomous systems built around language models, memory, planning, tools, and action execution. They also characterize agent protocols as standardized communication frameworks designed to improve interoperability and secure collaboration.

  • Definition and Characteristics of LLM agents: LLM agents combine language processing with autonomous decision-making and operate in real-world environments rather than only generating text.
  • Definition and Characteristics of LLM agents: Core agent components include a foundation model, short- and long-term memory, planning, tool use, and action execution.
  • Definition and Characteristics of LLM agents: Planning enables agents to decompose complex tasks into smaller subtasks and supports strategic problem solving and decision-process interpretability.
  • Definition and Developments of Agent Protocols: Agent protocols define rules, formats, and procedures for structured communication among agents and between agents and external systems.
  • Definition and Developments of Agent Protocols: Protocols aim to provide interoperability, standardized interactions, extensibility, security, and governance across heterogeneous agent systems.
  • Definition and Developments of Agent Protocols: The landscape includes model-centric MCP, enterprise-focused A2A, and open network protocols such as ANP.

3 Protocol Taxonomy

The paper classifies agent protocols along two dimensions: object orientation and application scenario. These dimensions distinguish context-oriented from inter-agent protocols and general-purpose from domain-specific protocols.

  • Protocols are classified by object orientation as either context-oriented or inter-agent.
  • Protocols are also categorized by application scenario as general-purpose or domain-specific.

3.1 Context-Oriented Protocols

Context-oriented protocols standardize how LLM agents acquire context from external data, tools, and services, addressing fragmented invocation interfaces. MCP exemplifies this approach through a client-server architecture that supports interoperability, scalability, and privacy.

  • LLM agents autonomously select and invoke external tools to obtain context needed for complex queries and goals.
  • Proprietary tool-use standards and provider-specific interfaces fragment invocation methods, increasing customization burdens for users and developers.
  • Context-oriented protocols provide standardized context acquisition, with general-purpose protocols supporting broad entities and domain-specific protocols optimizing particular use cases.
  • General-Purpose Protocols: MCP connects LLM agents to external data, tools, and services through a universal, open context-oriented protocol.
  • General-Purpose Protocols: MCP uses Host, Client, Server, and Resource components to coordinate context requests between agents and external resources.
  • General-Purpose Protocols: MCP decouples tool invocation from specific model and provider interfaces, improving interoperability, scalability, and cross-platform compatibility.
  • General-Purpose Protocols: MCP reduces data-leakage risks by having local clients construct and execute context requests and manage user authorization.
  • Domain-Specific Protocols: agents.json is an OpenAPI-based machine-readable contract format that declares AI-compatible interfaces, authentication schemes, and multi-step workflows.

3.2 Inter-Agent Protocols

Inter-agent protocols standardize communication among heterogeneous agents, addressing the lack of defined interaction methods in multi-agent systems. The section surveys protocol layers, architectures, and representative protocols for discovery, interoperability, security, and collaboration.

  • Inter-Agent Protocols address the absence of standardized interaction methods that constrains multi-agent system development.
  • General-purpose and domain-specific inter-agent protocols are distinguished according to their application scenarios.
  • Representative Protocols: Representative protocols target different interaction needs: ANP enables heterogeneous-domain interoperability, A2A supports enterprise collaboration, and AComP standardizes communication across organizations and frameworks.
  • Agent Network Protocol: ANP’s three core principles are interconnectivity, native machine-to-machine interfaces, and efficient collaboration through automatic organization and negotiation.
  • Agent Network Protocol: ANP uses decentralized identity and encrypted communication to support secure, trustless, end-to-end agent interaction.
  • Protocol Architecture: A layered architecture separates protocol negotiation, application-level discovery and capability description, and execution through authenticated requests and responses.

4 Protocol Evaluation and Comparison

The paper evaluates agent protocols through dimensions adapted to dynamic, semantic, and decentralized interactions rather than proposing a fixed benchmark. It emphasizes efficiency, scalability, security, reliability, and evolving protocol-specific costs.

  • Protocol evaluation considers interoperability, performance efficiency, reliability, scalability, security, evolvability, and simplicity.
  • Efficiency: Efficiency covers throughput, latency, handshake overhead, message size, semantic processing, task coordination, and token consumption.
  • Efficiency: TPS-N measures messages or tasks processed per second at concurrency level N, capturing capacity for concurrent agent interactions.
  • Scalability: Scalability measures performance as agent, tool, plugin, and communication-link counts grow, including degradation in latency and throughput.
  • Scalability: CNS evaluates capability negotiation through negotiation success rate and time as network size increases.
  • Security: Security evaluation covers identity authentication, encryption, integrity validation, authentication-mode diversity, access-control granularity, and contextual data desensitization.
  • Reliability: Persistent connections reduce repeated setup overhead, while UDR and MLR measure unexpected disconnections and messages failing to reach recipients.

4.5 Extensibility

Extensibility concerns whether agent protocols can evolve with new requirements while preserving compatibility and practical deployability. The section evaluates backward compatibility, adaptability, customization, operability, and observability.

  • Extensible Agent Protocols add or modify functionality without disrupting backward compatibility.
  • Backward Compatibility: Backward compatibility preserves normal interactions for existing clients after major protocol upgrades, measured by Upgrade Success Rate.
  • Flexibility and Adaptability: Flexible, modality-agnostic designs let developers add fields, semantics, and communication modalities for changing technologies and application scenarios.
  • Customization and Extension: Customization supports application-specific fields and plugins that add capabilities without affecting existing functionality.
  • Operability: Operability prioritizes concise specifications, framework agnosticism, and ease of implementation, operation, and maintenance.
  • Operability: Lightweight endpoints and predefined response models reduce code volume and simplify integration through Runs, Threads, and Store modules.
  • Operability: Documentation, SDKs, client libraries, monitoring, and debugging interfaces support deployment, configuration, observability, and diagnosis.

4.7 Interoperability

Interoperability enables agents across systems, platforms, browsers, and networks to communicate through standardized interfaces and formats. Protocol evolution illustrates trade-offs among broader functionality, security, performance, and collaborative complexity.

  • Agent Protocol interoperability abstracts implementation differences so heterogeneous agents can discover, communicate, and collaborate.
  • Interoperability Evaluation: Schema Compatibility Pass Rate measures communication without version conflicts or data-format issues.
  • Compatibility: Cross-system, cross-browser, cross-network, and cross-platform support maintains uniform agent communication across operating environments and changing network conditions.
  • Protocol Evolution: MCP v1.0-to-v1.2 evolution added HTTP Streaming and authentication, improving interoperability and security while potentially introducing new stream-latency factors.
  • Protocol Evolution: The MCP-to-ANP-to-A2A progression moves from structured context and tool integration toward decentralized identity and multilayer collaborative architectures.
  • Recommendations: Protocol development shifts from basic functionality toward complex collaboration, expanding ecosystem scalability and diversity.
  • Recommendations: Protocol selection should fit the application context, with MCP for external tools and data, ANP for cross-domain Internet collaboration, and A2A for broader inter-agent collaboration.

5 Use-Case Analysis

The case study contrasts four agent protocols on the same travel-planning task, showing distinct coordination patterns and application conditions. MCP centralizes tool use, A2A supports direct agent collaboration, ANP formalizes cross-domain exchanges, and Agora translates user intent into structured protocols.

  • MCP: MCP uses a central travel client to invoke flight, hotel, and weather services, aggregate responses, and generate the complete plan.External services function as tools in a strict star topology and do not interact with one another.
  • MCP: MCP offers simplicity and control but creates high dependencies, limited flexibility, and a potential bottleneck because all communication passes through the central agent.The central client must know every service and interface, making scaling or modification difficult.
  • A2A: A2A distributes intelligence across specialized agents that communicate directly, while a non-central travel planner primarily collects final results.Flight and activity agents can depend directly on the weather agent without routing every interaction through the planner.
  • A2A: A2A reduces unnecessary communication overhead and supports more complex collaboration patterns across organizations and multi-agent systems.Direct connections allow specialized agents to obtain needed information without central mediation for every exchange.
  • ANP: ANP formalizes structured requests and responses between agents across organizational and security boundaries, while retaining overall coordination by a travel planner.Its protocol-based interactions suit agents with distinct capabilities and well-defined interfaces.
  • Agora: Agora parses natural-language requests, generates formal protocols for service types, distributes them to specialized agents, and lets each agent respond to its assigned protocol.This three-stage separation shields domain-specific agents from natural-language interpretation and lets them focus on core competencies.
  • Cross-protocol comparison: Protocol applicability depends on desired agent autonomy, communication flexibility, interface standardization, and task complexity.The case study characterizes MCP as effective for well-defined workflows, A2A as flexible message-based collaboration, ANP as standardized delegation, and Agora as intent-to-protocol mediation.

6 Academic Outlook

The outlook identifies evaluation, privacy, group interaction, adaptability, layered architectures, collective intelligence, and agent-focused data infrastructure as major directions for protocol development. These directions move protocols beyond static pairwise exchanges toward modular, scalable, and ecosystem-level coordination.

  • Evaluation and Benchmarking: A unified benchmark remains underexplored, motivating evaluation frameworks that measure communication efficiency, robustness, adaptability, scalability, and task success.Standardized testbeds and diverse simulation environments are proposed to support fair comparisons across protocols.
  • Privacy-Preserving Protocols: Privacy-preserving protocols should enable information exchange in sensitive domains while minimizing exposure of internal states and personal data.Authorization may be managed using attributes such as agent role, task, or security clearance.
  • Agent Mesh Protocol: An Agent Mesh Protocol could support group-level communication with shared history and access, addressing inefficiencies in pairwise interaction as agent populations grow.Implementation requires group semantics, synchronized shared knowledge, message ordering, dynamic membership, and communication-overhead management.
  • Evolvable Protocols: Evolvable protocols would let agents retrieve, combine, and refine modular communication strategies for particular tasks, partners, or conditions.Agents could discover protocol variations or negotiation strategies that improve communication efficiency or task success over time.
  • Built-In Protocol Knowledge: Embedding protocol knowledge in model parameters could produce protocol-compliant behavior without inference-time instructions, but may reduce adaptability when protocols change.The supplied passage identifies adaptability as the main limitation of directly injecting protocol knowledge through training.
  • Layered Protocol Architectures: Layered architectures could separate transport and synchronization from semantic and task interactions, improving modularity and scalability across heterogeneous agents.They may also support dynamic layer selection and composition while integrating ethical, legal, and social constraints at higher layers.
  • Collective Intelligence and Scaling Laws: Future research may study how population size, communication topology, and protocol configuration jointly shape collective intelligence in interconnected agent populations.This direction builds on multi-agent systems, swarm intelligence, and complex adaptive networks.
  • Agent Data Networks: An Agent Data Network could provide structured, intent-driven, protocol-compliant information exchange optimized for autonomous agent communication and coordination.The proposed infrastructure would operate on top of the existing internet stack while differing from web interactions designed primarily for human interpretation.

7 Conclusion

The survey classifies and evaluates AI agent protocols to clarify their trade-offs and design considerations. It also points toward evolvable, privacy-aware, group-coordinated, and layered infrastructures for a more connected agent ecosystem.

  • Survey contribution: The survey systematically classifies agent protocols and evaluates efficiency, scalability, and security to help practitioners and researchers navigate the growing protocol ecosystem.The overview is intended as a practical reference for selecting and designing agent communication infrastructures.
  • Future direction: The paper highlights trade-offs in building reliable, efficient, and secure agent systems and envisions protocols that support dynamic coalitions, knowledge exchange, and collective intelligence.Future directions include evolvable, privacy-aware, group-coordinated protocols, layered architectures, and collective intelligence infrastructures.
Loading 2504.16736v3…