A Technical Taxonomy of LLM Agent Communication Protocols

As large language models (LLMs) advance and multi-agent systems aim to overcome the limits of standalone agents, robust communication protocols are becoming essential infrastructure for distributed agent networks. Nonetheless, the fragmented protocol landscape presents a significant interoperability challenge. This study develops a technical taxonomy to classify and analyze LLM agent communication protocols. Following an established iterative method, we defined the taxonomy’s purpose, meta-characteristic, and ending conditions, then performed five iterations, three empirical-to-conceptual and two conceptual-toempirical, on nine actively maintained open-source protocols with demonstrable adoption. The taxonomy comprises five dimensions: counterparty, payload, interaction state, discovery mechanism, and schema flexibility. Classification reveals recurring architectural patterns: all sampled agent-to-agent protocols combine hybrid payloads with session-state persistence; most protocols support multiple predefined schemas, and two negotiate schemas at runtime, indicating a trend toward schema flexibility; decentralized discovery remains rare. Analysis suggests short-term convergence pressure toward protocols unifying agent-to-agent and agent-to-context (tool and data) communication.

Introduction. Recently, large language model (LLM) based agents have attracted significant attention in AI research [1–3]. LLMs incorporate extensive world knowledge [4] along with advanced reasoning and planning capabilities [5–7]. Embedded as the agent’s core and equipped with memory, sensors, and actuators, the LLM grants the agent the capacity to interact dynamically with its environment and to perform efficient multistep reasoning, thereby enabling it to solve complex real-world tasks [8–10]. Building on single LLM agents, a promising next frontier involves multiple cooperating agents. Such multi-agent systems (MAS) consist of specialized agents that communicate, collaborate, and debate with one another [11, 12]. Rooted in the principles of collective intelligence [13], this approach further enhances the system’s problem-solving capabilities, yielding outcomes superior to those achievable by any individual agent [11, 14, 15]. Here, one detail must not be overlooked: without a communication mechanism, no collective intelligence can arise.

Discussion / Conclusion. Applying Marro’s [26] Agent Communication Trilemma, introduced in Section 3, to our taxonomy reveals stark architectural trade-offs. The trilemma dictates that a protocol cannot simultaneously maximize versatility, efficiency, and portability. This trade-off is highly visible at the extremes of our Schema Flexibility and Payload dimensions. Protocols designed primarily for context-interaction, such as MCP, maximize portability and efficiency by enforcing rigid schemas, stateless interactions, and strictly structured data payloads. This strictness eliminates token-heavy negotiation, making them ideal for predictable, high-throughput agent-to-tool invocations. Conversely, protocols with evolving schemas, such as Agora and ANP, maximize versatility to facilitate dynamic, open-ended multi-agent debates.