Countless AI agents are performing countless functions, but are they talking to each other? It is time to think about how to connect them, both within and across organizations.

An open-source collective called AGNTCY has been formed, intended to provide the protocols and standards for an open and interoperable "internet of agents" that crosses organizational and industry boundaries.

That is the word from Tushar Agrawal, senior director of AI at Cisco, whose remarks were delivered at Fabrix.AI's recent Agentic Demo Day. Participants and contributors to the AGNTCY collaboration include Cisco, LangChain, Galileo, Fabrix.AI, MongoDB, and Boomi.

Agent connectivity

The case being made for agent connectivity may sound familiar to anyone who has been involved in building and managing microservices or service-oriented architecture -- publishing or subscribing and orchestrating standardized services to deliver a particular function needed at the moment.

The idea is to enable an ensemble of agents "that collaborate and get tasks done," Agrawal said. "How do we enable that these agents can talk to each other, can understand each other, can interact securely, can interact efficiently in a low-latency way? That's those are some of the challenges and standardization work that agency is looking to do."

Such an internet of agents "layers on top of the cloud internet," he explained. "Just like in the internet era, where there were standardized protocols -- Internet Protocol (IP), Domain Name System (DNS), Border Gateway Protocol (BGP) -- which enabled these services to collaborate with each other. We think that there are going to be a similar set of protocols and similar set of standards that have to be established so that the agentic apps and intra-agent and inter-agent collaboration can happen."

The collaboration of AI agents, "which are cross-domain, cross-vertical, cross-purpose in a seamless fashion within organizations and across organizations over the internet is what we are terming as the internet of agents architecture," he stated.

The components

There are three main components, developed as protocols through AGNTCY, to leveraging the internet of agents, Agrawal elaborated:

• Discover: AGNTCY provides an "agent directory, based on the OCI open container registry framework," he detailed. "These directories are designed in a way that they can sync with each other, so you can have multiple directories within an organization." The result is a "DNS-like service which can enable agent publishers to publish agents, and also enables agent consumers to discover those agents, and resolve to the place where those agents are actually stored and understand all the dependencies." The consortium offers an Open Agentic Schema Framework, which serves as a schema in which every agent developer can publish their agent. "It is a standard way of describing the capabilities of the agent, so that it can be discovered, it can be understood in terms of its capabilities, can be selected, it can be republished."
• Compose: Once discovered, multiple agents need to be brought together, said Agrawal. "How do you really get these cross-framework platform agents to operate together, and what does that set of components look like?" Agrawal said. "How do you deploy them? What does the runtime aspect of that look like? How do you make sure that they are doing what they're supposed to be doing -- and they're not doing what they're not supposed to be doing?"
• Deploy and run: "There is a good play here for something like an agent gateway protocol or an agent messaging protocol, which enables this publisher/subscriber way for consuming the information that is being shared between various agents and the tools. You have the MCP, which is one of the ways in which foundation models or agents can get data or tools or prompt access." This provides for a "publisher/subscriber highly secure messaging bus, so that all of this traffic is sent securely with low latency -- despite sort of being highly multimodal in a very effective way between all these various frameworks and agents together."

The lessons learned from the development of the internet provide an example of how multiple AI agents can be aligned to support business processes. "Today, you don't have to even think about standing up a website and making sure that people can get to it -- and all of the translation, routing, and all of that is taken care of in a standard way," said Agrawal. "We envision the same to happen with agents, where you have these agentic applications and they are published within an org boundary or in a global boundary."

Get the morning's top stories in your inbox each day with our Tech Today newsletter.