The Linux Foundation today at the Open Source Software Summit revealed that the Agent2Agent (A2A) protocol created by Google will now being advanced under the auspices of the consortium.
Launched earlier this year, the A2A protocol defines a way for artificial intelligence (AI) agents to communicate with each other that is now supported by more than 100 IT vendors, including Amazon Web Services (AWS), Microsoft, Cisco, Salesforce, ServiceNow and SAP.
At its core, the A2A protocol creates an AgentCard, a JSON metadata document that makes an AI agent discoverable using a URL in addition to describing its capabilities. Whenever a client sends a message to an agent, the agent might determine that fulfilling the request requires a stateful task to be completed, with each task having a unique ID defined by the agent.
Mike Smith, a staff software engineer for Google, told conference attendees that the A2A protocol has also been modified to make it simpler to add custom extensions to the core specification. Additionally, the A2A community will also soon be focusing on making it simpler to apply unique identities to AI agents to improve overall governance.
In general, the A2A protocol is squarely focused on interactions between AI agents, whereas the Model Context Protocol (MCP) originally developed by Anthropic is designed to make it simpler for AI agents to ingest data and interact with existing tools. As such, the A2A protocol and MCP are a complementary set of initiatives, said Smith. “Both should exist together,” he says.
Given the use cases, MCP is being more widely adopted at the moment, but both initiatives are still relatively new so many of the use cases remain experimental. The A2A initiative is a similarly nascent project.
However, they both make it clear that a foundation upon which multiple AI agents can be orchestrated. Less clear, however, is to what degree additional initiatives might be required to achieve that ultimate goal.
The one thing that is apparent is that AI agents will drive a massive wave of business process reengineering. A Futurum Group report projects AI agents over the course of the next three years will drive $6 trillion worth of economic value by 2028.
There are, of course, a wide range of issues that will need to be addressed to realize that promise, not the least of which is defining the frameworks required to govern and secure AI agents. Unlike previous innovations, AI agents are designed to function autonomously. While that will undoubtedly boost productivity, there remains a possibility that an AI agent will go rogue in the sense that it performs unexpected tasks. Additionally, AI agents also present a rich target for cybercriminals that will now be able to not just steal data but also commandeer entire workflows.
None of that means that AI agents won’t be pervasively adopted over the next few years. However, as is the case with most innovative technologies, the level of impact they might have in the short term may be overstated, while in the long term it may be greatly underestimated.
