A Technical Taxonomy of LLM Agent Communication Protocols
This study presents a technical taxonomy of LLM agent communication protocols based on a five-dimensional analysis of nine open-source systems, revealing current architectural patterns and predicting a future evolution toward a federated, layered protocol stack rather than a single universal standard.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where artificial intelligence is no longer a single, solitary mind answering questions, but a bustling society of specialized workers. Some of these digital workers are experts at finding information, others at writing code, and others at making complex plans. To solve a difficult problem, they must talk to one another, share their findings, and coordinate their actions. This is the promise of multi-agent systems: a group of intelligent programs working together to achieve what none could do alone. However, for this society to function, these digital workers need a common language. Just as humans need to agree on grammar and vocabulary to hold a conversation, these software agents need a set of agreed-upon rules to exchange information. Without these rules, a worker built by one company might not be able to understand a worker built by another, leaving them isolated in their own silos.
A team of researchers at the Technical University of Munich has stepped into this emerging field to map out the current landscape of these communication rules. They set out to understand how different systems are currently trying to connect these digital agents. The researchers did not invent a new language; instead, they acted as cartographers, drawing a detailed map of the nine most active and open communication protocols that are currently being used and developed by the community. By examining these nine systems closely, they created a new way to classify them, revealing that while there is a lot of activity, the field is still fragmented and moving toward a specific future structure.
The researchers began by selecting nine specific protocols that were actively maintained and openly available. They looked at systems designed to let agents talk to other agents, as well as systems designed to let agents talk to tools, databases, or other external services. They then analyzed each one through five specific lenses. First, they looked at who the agent was talking to: was it another agent, a tool, or both? Second, they examined the content of the messages: were they just text, just structured data like numbers and files, or a mix of both? Third, they checked if the conversation kept a memory of what had happened before, or if every message was treated as a fresh start. Fourth, they investigated how the agents found each other in the first place: did they need a central directory to look up addresses, or could they find each other directly? Finally, they looked at the flexibility of the conversation rules: were the rules fixed in stone, could the agents choose from a few options, or could they invent new rules as they spoke?
The analysis revealed a clear pattern in how these systems are built. Every protocol designed for agents to talk to other agents supports a "session state," meaning the conversation remembers its history, which is essential for complex, multi-step tasks. None of these agent-to-agent systems rely solely on simple text; they all handle a mix of conversation and structured data. The researchers also found a strong trend toward flexibility. Most protocols allow agents to choose from several pre-defined ways to structure their messages, and two of the protocols go even further, allowing the agents to negotiate and change the structure of their conversation while it is happening. This suggests that as these systems become more advanced, the ability to adapt the rules on the fly is becoming a standard requirement.
However, the study also highlighted a significant gap. While many protocols allow agents to talk to each other or to tools, very few support a truly decentralized way for agents to find one another. Most rely on a central registry or a fixed address, much like a phone book or a pre-written contact list. This centralization could become a bottleneck if the number of agents grows to the scale of the internet. The researchers noted that only one of the nine protocols they studied offered a fully decentralized method for discovery, where agents find each other without a central authority.
The most important finding of the study is a prediction about the future of this technology. The researchers argue that it is unlikely a single, all-purpose protocol will emerge to rule them all, capable of handling everything from simple tool calls to complex, long debates. Instead, they suggest the field is moving toward a layered system, similar to how the internet itself is built. In this future, different protocols would handle different layers of communication. Some would be lightweight and rigid, perfect for quickly asking a tool to perform a simple task. Others would be flexible and memory-rich, designed for the complex, evolving conversations between multiple agents. This layered approach would allow the "Internet of Agents" to be both efficient and versatile, using the right tool for the right job at every level of interaction.
The study concludes by pointing out a missing piece in the current puzzle: privacy and safety. None of the nine protocols examined included built-in mechanisms to enforce strict privacy rules or to check if an agent is allowed to perform a specific action. As these systems move into real-world applications like healthcare or finance, the researchers suggest that a new layer of governance will be needed to ensure these digital conversations remain secure and compliant with laws. For now, the field is in a period of rapid experimentation, but the path forward seems clear: a diverse, layered ecosystem of protocols working together, rather than a single standard trying to do everything.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.