Given that the point of an AI agent is to complete tasks, not just answer questions, it’s no surprise there’s a range of options for connecting agents to the tools and services that allow them to take actions. But with the launch this summer of yet another open agent protocol, Agentic Resource Discovery, it might seem like everyone is coming up with their own standard.
MCP came from Anthropic, A2A from Google, and the ARD spec authors are from Microsoft, Google and Hugging Face, with contributions from Cisco, Databricks, GitHub, GoDaddy, Nvidia, Salesforce, ServiceNow and Snowflake. That free-for-all makes sense: controlling how services can be found can also influences which services are findable easily (or possibly at all).
There’s also ACP (IBM’s Agent Communication Protocol, not to be confused with the Agent Client Protocol that connects coding agents to IDEs, or with the Agentic Commerce Protocol from OpenAI and Stripe), AGNTCY (a Linux Foundation project originally from Cisco for agent discovery and identity that builds on the Open Agent Schema Framework), the Agent Network Protocol stack, and AGTP, an Agent Transport Protocol proposal to move agent traffic away from the HTTP that all the other protocols rely on.
“It's early, so there's a land grab of different organisations trying to put their stake in the ground with some protocol,” Forrester principal analyst David Mooter told The Stack. “They’re all tackling the problems [in] different ways.”