MCP vs ACP vs A2A: Comparing Agent Protocols - Laurie Voss (AI column)
Ironic. speech Laurie Voss, VP Developer Relations in LlamaIndex, where he 15 He spoke about the main protocols of interagency interaction. He originally wanted to talk about it. 3 The main protocol, but in the course of preparation found more 11:) And in the course of his speech, he joked that he would add WTF and LOL to the list of protocols for nothing. There are protocols with such names.
The main ideas are presented below.
1. Creating protocols is now fashionable In a month of research, the scope of the report increased 3 before 14 It seems that almost everyone is working on agency protocols. 2. Two types of protocols
- Context-orientedProvide the context of the LLM (MCP, agents.json)
- Inter-agentInteraction between agents (others) 3. MCP - De facto standard Model Context Protocol Anthropic dominates due to its early focus on a specific problem. 4. A2A from Google Agent2Agent Protocol Google focuses on asynchronous interaction and long-term tasks. 5. Two ACP protocols
- ACP Connect from AgntcyIncludes Cisco, LangChain, LlamaIndex, Galileo and Glean with an integrated agent registry
- ACP Communication from IBM - fork MCP, developing towards A2A 6. Agora is a self-updating protocol Agora It's an interesting approach where it starts with natural language and then it allows participants to update the protocol during interaction. 7. Protocol from the blockchain guys AITP NEAR Protocol focuses too much on the cost of interactions. It’s funny that the speaker joked that “blockchain hasn’t solved a single problem yet, so I don’t think it will solve this one.” 8. LMOS almost from IBM The Eclipse Foundation is creating LMOSIn parallel with its own ACP. There was a joke about the Eclipse Foundation being a disguised IBM building the same thing that IBM is building. And if that sounds strange to you, you haven't talked much to IBM.) 9. Lack of key components According to the author, all protocols lack a centralized registry, an authorization system and a reputation system. 10. MCP as the foundation of the future The bottom line is that MCP has gained attention and can expand to solve inter-agency problems without adding new protocols.
P.S. This agent topic is a good addition to the story from yesterday’s report on AI implementation in SDLC. told CTO Conf X.
#AI #Agents #Engineering #DistributedSystems