Glossary · Protocols and standards
Agora protocol
A research meta-protocol for LLM agents: they agree on plain-text protocol documents identified by hash, and fall back to natural language.
The Agora protocol is a meta-protocol for networks of LLM-powered agents in which agents use natural language for rare exchanges, standard routines for frequent ones, and negotiated protocol documents for the cases in between.
Origin. Samuele Marro and six co-authors introduced Agora in an October 2024 paper (arXiv 2410.11905). The paper frames the problem as the Agent Communication Trilemma: communication across a large agent network should be versatile, efficient and portable, and large networks struggle to achieve all three at once. Agora’s answer is to let agents agree on a structured format when an exchange repeats often enough, and to keep natural language as the fallback.
How it works. A sender POSTs a JSON document to the receiver’s Agora endpoint. The protocolHash field names the protocol in use, and null means plain natural language. The documentation’s first example:
{"protocolHash": null, "body": "Hello! What's the weather forecast for London on April 25th, 2025?"}
The receiver answers with a status of success or failure and a body. A protocol document is a text file: YAML metadata with a name, a description and a multiround flag, followed by a plain-language description of the messages. Agents identify a protocol by the SHA-1 hash of that file. A receiver can list the protocols it supports at a /wellknown endpoint, and it treats a partly supported protocol as unsupported. Multi-round exchanges get a conversation ID, and one conversation keeps one protocol throughout. The specification does not fix how agents negotiate a protocol, or whether an LLM or generated code handles a message.
Status. The documentation is at version 0.2, the code is MIT-licensed, and the project describes itself as fully decentralized, with no central node. Its public GitHub repositories show no updates since March 2025.
Neighbouring terms. A2A takes the opposite approach: one fixed message and task model that every implementation shares. ANP includes a draft meta-protocol layer for similar negotiation.