In April, Google announced the Agent2Agent (A2A) protocol at its Cloud Next conference. The protocol aims to make it easier for AI agents to talk to each other, no matter what framework they were built with. Since everything in AI moves at lightning speed these days, at the Open Source Summit in Denver, Google today announced that it has donated the protocol to the Linux Foundation and moved it to a new GitHub repository.
Among the companies joining Google in what is unsurprisingly called the Agent2Agent project are AWS, Cisco, Salesforce, SAP and ServiceNow. The project will shepherd the future of the protocol itself, as well as the SDKs, NPM packages and other developer tooling.
As with similar projects, the objectives here are to establish an open standard for agent interoperability, foster an ecosystem around that standard and — based on the standard Linux Foundation framework — ensure neutral governance.
“When we set out to design Agent2Agent, we knew we needed a truly open protocol that would work for anyone, be it individual developers all the way up to larger enterprises,” Mike Smith, a staff software engineer at Google who has been driving the project internally, said during his Open Source Summit keynote. “And we knew it needed to be independent and vendor-agnostic […]. And so we had this dream of like, what if we can join the Linux Foundation? I’ve seen the success of other Linux Foundation projects.”
But What About MCP?
Image credit: The New Stack/Frederic Lardinois.
During his keynote, Smith also addressed a common misconception about A2A: it’s not meant to compete with Anthropic’s Model Context Protocol (MCP). While MCP has been the breakout hit of agentic protocols, the idea behind MCP is to connect agents to tools and data sources. A2A is all about connecting agents to each other. That includes, among other things, making them discoverable and helping them advertise what capabilities they have so that a crew of agents can more effectively work together to solve a given problem.
“MCP makes a lot of sense. It’s definitely taken off as a protocol for interacting with tools,” Smith said. “Agents are a little bit weirder. They’re a little bit ambiguous. There’s more of a negotiation process that happens when you interact with an agent.” You may ask an agent to do something, Smith explained, and that agent may ask you for more information, or let you know that it can do part of a given task, but not all of it.
MCP and A2A complement each other, but as Smith also stressed, it’s still early days for these protocols and he expects to see more of them in the future.
Image credit: The New Stack/Frederic Lardinois.