What do we assume about artificial intelligence that might be holding us back from getting more centered on how to navigate the future? What could help us with collaborative models that put our societies in a better place to integrate something that’s been compared to electricity, or fire, in its global impact?

As the agentic future approaches, we see more of what AI is capable of. In the earlier days, there was the “Watson” era: the chatbot teaching us that it can recall any fact from the haystack of a knowledge base, and beat humans at calculation and deterministic thinking.

Now, agents are taking us beyond this simple premise, into the uncharted waters where the AI is actually going to do things for us, begging the question: what will it do, and how will we work together with non-human partners, assistants and collaborators?

So the assumptions themselves are valuable. They help us to put together a road map – where we want to go, where we don’t want to go, and ultimately, where we are going.

At this year’s April event for Imagination in Action, I sat in on a panel where the group explored some of the assumptions that might get in the way of clear-eyed assessment, or ideas that may be a little off the mark for figuring out a converged answer to AI development. (Disclaimer: April’s IIA event is an annual conference that I help to facilitate.)

Panelist Ayush Chopra, an MIT PhD candidate, started by taking on the very core idea that an AI agent, by itself, constitutes an “intelligent” force that has the answers within it.

“A lot of the agents today are built with this fundamental assumption that all intelligence exists in the agent, that the agent is intelligent enough that it'll figure everything out on its own, but when everybody's intelligent, basically nobody is, so I think like the fundamental assumption for the future is that intelligence is in the interaction, so what you'll start to see is: you need to learn to design the correct interaction protocols, rather than just assuming all intelligence sits in the agent.”

Cutting at the notion of infallibility with AI, panelist Abhishek Mehta, CEO of Tresada, described a number of layers inherent in the technology, many of which, he suggested, are failure-prone.

“In our world today, data is a very centralized thesis, whether it's application based, whether it's storage based, or whether it's compute-based,” he said. “We all assume data, and the management of it, should be centralized. In a world of agents, it doesn't work, because you cannot wait for a mothership. I don't think we'd be comfortable with a mothership, but we can't wait for a mothership to provide the right signals needed to create an intelligent system that can make decisions when they need to be made, just like humans do. So, in my opinion, with the idea of a tech stack, with chips at the bottom and agents at the top, everything in the middle is up for grabs.”

Mehta called for a different approach to AI agents, citing the work of Tim Berners-Lee in the design of the early internet with HTTP, HTML and URLs.

“I think the fundamental building blocks of an agentic ecosystem must be worked on in a community,” he said, expressing the opinion that “too much greed” will make it hard for the open source movement to make inroads. “That is not what's going on right now.”

Panelist Manuela Veloso, professor emeritus at Carnegie Mellon, took on the assumption that agents will easily be able to communicate seamlessly, trading information like adolescents might trade baseball cards, or pokemon cards, or whatever’s hot today. She gave this example:

“I might ask, what's the address of the CEO, and an agent replies (with) the name of the street, but doesn't give the zip code, doesn't give the state, doesn't give anything. And then I meant address, complete address - when an agent asks something to another agent, I think what breaks is the is this assumption that what I ask for is understood at the level of detail that I need, so that's a challenge, to have multiple computer entities talking with each other, and assuming that we live all in the same space and level of detail.”

As for MIT associate professor Phillip Isola, his musing considered whether the future AI agents might be more inspired by biological design, less deterministic, and more “organoid.”

“One possibility is, this future will look more like economies, and markets, and social policies, and less like networks and computer systems, computer science, a lot of the ways that we build the tools right now,” he said. “The LLMs right now look very much like CS. They follow some protocol, and you have this MCP thing, and so on and so forth. But maybe a better model for the future of these agents is to think of them like organisms, not like programs.”

Recursive AI On the Horizon

“Very soon,” Mehta suggested, “agents will build agents, which will have less fidelity, and when agents who build agents build more agents, the agent will evolve quite rapidly, and drive a lot more tactical development of what agents will do, to automate 70% of the world's economy.”

Chopra agreed, citing AI’s approach of something called the “Dunbar ceiling,” a socialization principle, and explaining a recursive process this way:

“In our lab, we spent the last five years just studying large multi-agent systems, and for many years, human societies were a big motivation, so we used to use agents to model disease outbreaks, labor markets, social networks, supply chains, that was the bulk of the focus of what we would do to study large models of societies, and how agents sort of self-organize at scale.”

“How do you build repetitions of these systems and these agents?” asked panelist Anil Sharma of Tata Consultancy Services, referring to scaling. “Who backs them up? There are still a lot of questions to be addressed. The good thing is: we have a lot of knowledge of how things have worked today, and some of those ideas are potentially scalable, scalable to even internal agents.”

“They will have limitations,” Veloso added, suggesting that the really impressive part of the agents will be in the interactions with humans.

Those are some thoughts on what we might see with AI agents moving forward in the next few years, and how they may challenge some of the assumptions that so many of us have. As for the open source movement – we’ll see.