As large language models (LLM) agents receive traction in enterprises and research ecosystems, a fundamental difference has come out: communication. While agents today can autonomally cause, plan and work, their ability to coordinate with other agents or interface with external devices remains forced by the absence of standardized protocols. This communication bottleneck not only cuts the agent landscape, but also limits scalability, interpreting and emergence of collaborative AI systems.
A recent survey conducted by researchers from Shanghai Jiao Tong University and ANP community provides the first comprehensive classification and evaluation of protocols for AI agents. The work introduces a royal classification plan, discovering the existing protocol framework, and scalable, safe and intelligent agent underlines future directions for ecosystems.
Communication problem in modern AI agents
The deployment of LLM agents has carried forward the development of mechanisms that enable them to interact with each other or with external resources. In practice, most agent interaction rely on ad hoc API or brittle function-jolling paradigms-in which generality, safety guarantee, and cross-vynder lack of compatibility.
The issue is in line with the early days of the Internet, where the absence of general transport and application-layer protocol prevented spontaneous information exchange. The way TCP/IP and HTTP covered global connectivity, standard protocols for AI agents are ready to serve as the backbone of future “Internet of agents”.
A framework for agent protocol: reference versus cooperation
The author proposes a two-dimensional classification system that portrays the agent protocol with two axes:
- Reference-oriented vs. inter-agent protocol
- Reference-oriented protocol How the government interacts with external data, equipment or API.
- Inter-agent protocol In many agents enable colleague to colleague communication, work delegation and coordination.
- General-purpose vs. domain-specific protocol
- General objective protocol Miscellaneous environment and agents are designed to work in types.
- Domain specific protocol Human-agent is adapted for special applications such as dialogues, robotics or IOT systems.
This classification helps explain the design business during flexibility, performance and expertise.
Major protocols and their design principles
1. Model Reference Protocol (MCP) , anthropic
The MCP is a general-reference reference-oriented protocol that facilitates structured interaction between LLM agents and external resources. Its architecture reduces argument (host agents) from an execution (customers and servers), increases safety and scalability. In particular, MCP reduces privacy risks by ensuring that sensitive user data is processed locally, rather than being embedded directly into LLM-Janit function calls.
2. Agent-to-agent protocol (A2A) , Google
Designed for safe and persuasion, A2A enables agents to exchange tasks and artifacts in enterprise settings. It emphasizes modularity, multimodal support (eg, files, streams), and opaque execution, which preserves IP competent by enabling interoperability. Protocol defines standardized institutions Agent card, WorkAnd Artifacts For strong workflow orchestation.
3. Agency network protocol , open source
ANP implements a decentralized, web-scale agent network. The created decentralized identity (DID) and the cementic meta -otocol layers, ANP facilitates reliable, encrypted communication between agents in the heterogeneous domains. It introduces layered abstraction for discovery, interaction and performance – an open “internet of agents” depict themselves as a foundation.
Performance Metrix: A overall assessment structure
To assess the strength of the protocol, the survey introduces a comprehensive outline based on the seven assessment criteria:
- Capacity – Throopoot, delay and resource usage (eg, token in LLM)
- Adiposity – Support for increasing agents, dense communication and dynamic work allocation
- Security -The Fine-Grand Authentication, Access Control and Reference Dysensitization
- Reliability – Strong message distribution, flow control and connection firmness
- Tanana – The ability to develop without breaking compatibility
- Operatingness Ease of tanati, observation and stage-unknown implementation
- Introopreciability Cross-system compatibility in language, platforms and vendors
The structure refers to both classical network protocol principles and agent-specific challenges such as cementic coordination and multi-turn workflows.
To emerging collective intelligence
Protocol lies in capacity for one of the most compelling arguments for standardization collective intelligenceBy aligning communication strategies and abilities, agents can build dynamic aligning to solve complex functions – which put robotics or modular cognitive systems into herds. Like protocol Agora Take it forward by enabling agents to interact and optimize new protocols in real time using LLM-related routines and structured documents.
Similarly, like a protocol Loca Embed moral logic and identification management in the communication layer, ensuring that agents ecological mechanisms can develop transparent and safely.
The Road Ever: Static Interface to Adaptive Protocol
Looking forward, the author prepare a three -phase outline in protocol development:
- Short term: Infection from rigorous function calls dynamic, for evolvable protocols.
- Mid -term: The agent from the rule-based API enables self-organization and interaction in the ecosystem.
- Long -term: The emergence of layered infrastructure that supports privacy-conservation, collaborative and intelligent agent networks.
These trends indicate departure from traditional software design to a more flexible, agent-element computing paradigm.
conclusion
The future of AI will not be shaped only by model architecture or training data – it will be shaped by dialogue, coordination and learning ways to each other. Protocols are not just technical specifications; They are connective tissues of intelligent systems. Formally to these communication layers, we unlock the possibility of a decentralized, safe and interopeable network of agents – an architecture that is capable of scaling beyond any single model or framework capabilities.
See model on paperAlso, don’t forget to follow us Twitter And join us Wire And LinkedIn GROUPDon’t forget to join us 90K+ ML Subredit,
[Register Now] Minicon Virtual Conference on Agentic AI: Free Registration + attendance Certificate + 4 hours small incident (May 21, 9 am- 1 PM PST)
Sana Hasan, a counseling intern and double degree student at Marktekpost in IIT Madras, is emotional about implementing technology and AI to resolve real -world challenges. With a keen interest in solving practical problems, he brings a new approach to the intersection of AI and real -life solutions.
