Agentic AI communications: Identifying the standards we need

When it comes to standards work around agentic AI, we’re at an exciting threshold. As more tools emerge, we’re seeing the amazing things it can accomplish. Now, we’re trying to figure out what parts of it need to be standardized.

There’s lots of work going on in this space, particularly at the IETF. But first, I want to identify a few of the processes and problems that might merit standardization when it comes to multi-vendor interoperability of applications over the Internet. (Remember, these are just my opinions on the subject.) 

During IETF 124 in Montreal, Jonathan Rosenberg and I held a side meeting on this topic. You can view the strawman charter created for a potential working group. 

With agentic AI (i.e. systems that allow AI agents to act on a human’s behalf), there’s a lot of communication going on: There’s human-to-agent, agent-to-tools, agent-to-agent. The human-to-agent communication is often very specific to that particular agent and how it wants to interact, so I don’t see a need for too much standardization there. But I think agent-to-agent or agent-to-tool communication will need it, because this often involves agents from one organization interacting with agents from another organization. 

Let’s say you’re planning your week at IETF 125, and you ask a chatbot (typing) to tell you the best restaurants in Shenzhen, and that chatbot responds with a list. As you continue to communicate with the chatbot to narrow down where you want to go, it will probably be easier to just switch to talking (audio) or perhaps look at photos from the restaurants.