New arXiv paper classifies LLM agent communication protocols with 5 dimensions

A new paper on arXiv (2606.19135) by Linus Sander and colleagues at TU Munich presents a systematic taxonomy of LLM agent communication protocols. As multi-agent systems grow more complex, robust and interoperable protocols become critical infrastructure. The researchers developed an iterative classification framework with five dimensions: counterparty (agent-to-agent vs. agent-to-context), payload structure, interaction state management, discovery mechanism, and schema flexibility. They applied this to nine actively maintained open-source protocols with demonstrated adoption.

The analysis reveals several recurring patterns. All sampled agent-to-agent protocols combine hybrid payloads (text, structured data, and code) with session-state persistence. Most support multiple predefined schemas, and two protocols even negotiate schemas at runtime, indicating a clear trend toward greater flexibility. However, decentralized discovery remains rare. The authors predict short-term convergence pressure toward protocols that unify agent-to-agent and agent-to-tool/data communication. In the long term, no single protocol is likely to maximize versatility, efficiency, and portability simultaneously—instead, the field will likely evolve toward a federated, layered protocol stack. The taxonomy provides practical guidance for protocol selection and highlights open research gaps in privacy and policy enforcement.