As we move into a time of agentic AI, and there’s a lot of consensus around this, we start to think about how the environment will look: how will the AI agents talk to each other? What kinds of self-assembly will be the norm?

“A new organizational blueprint is emerging,” writes Alex Fleck at Microsoft. “It blends machine intelligence with human judgment to create systems that are AI-operated, but human-led.”

That word, “led,” carries a lot of weight. How, exactly, will humans be managing agent swarms? And how much of a human in the loop will remain in the average process?

I was at an MIT event in April where I heard a panel talking about these questions. At the Imagination in Action conference, Debo Dutta, Chief AI Officer at Nutanix, interviewed Antonio Figueiredo, senior director at Salesforce, Sam Sharaf, a project lead at Google, Partha Madhira, an engineer at Kyndryl, and Rudina Seseri, Founder and Managing Partner of Glasswing Ventures.

Dutta had his own concept of agent swarm systems, one drawn from pop culture.

“This is my cartoon version,” he said. “I love cartoons and the ‘Despicable Me’ series, so I think of (agents) as ‘minions.’ So if you don't completely put the right guardrail and governance around the minions, and the minions can be owned by different gurus of the world, talking to each other, I think it'll have a big societal impact, which way we choose, and how we build the structure of minions and hierarchies of minions, it's going to change our society quite a bit.”

Will Agent Connections be Proprietary?

Going down the line, Dutta asked each panel member to explain whether they see the agentic future as internal teams of agents under different corporate banners, or a more free and open “internet of agents” where common protocols will allow free interplay.

The consensus was that the internet of agents will be the future, with some of the panels providing disclaimers and arguing that it will take a while to get to that point, with some growing pains along the way.

“I'm in the ‘internet camp,’” Figueiredo said, “but I think the internet is going to stay a little bit, probably the next 18 months while companies are going through this initial experiment, initial work that they are doing today, and I think that's fine, that's helpful. It helps them to flex those muscles, to understand the existing bonded use cases, and workflow to get experience in terms of governance, in terms of capabilities.”

“You need agents to talk across organizational boundaries,” Madhira said, “you need agents to be exposed as service providers, because business models are also changing for AI companies, and for companies which are going to become AI companies in the future.”

“There are employee agents which can do workflows and all that,” Sharaf added, “but I think when it comes to the Internet of agents, we take the complexity and the scale to a different level.”

Seseri had a sort of blended outlook.

“I do believe that for the most part, this is going to be an internet of agents,” she said, characterizing it as a “winner takes most” scenario, “but I don't think it will be solely internet. I think there will be the intranet piece, and if I go along with the same analogy around the cloud, yes, you have the public clouds, but you also have your virtual private clouds, and they coexist, and we know one is significantly bigger than the other, but there is a role for each to play.”

Inspecting the Layer Cake

The next question that Dutta asked the panel was about the most important layers of the tech stack.

Figueiredo noted that observability in the process is important.

Sharaf cited the value of orchestration, trust, and governance.

“If the agent cannot tell you what it did, who it talked to, and why, it's not an enterprise agent, it's a liability,” he said.

Madhira used a different lens to make a similar point.

“If you don't have the right kind of telemetry data coming out of agents, then this just becomes distributed uncertainty,” he said.

“In a multi-agent system, identity is the real key,” Seseri said. “Identity is where security and cybersecurity and governance break down, right now they're the points of vulnerability, and in fact, as you think about the handoff from one agent to the next agent, without having that clear identity, it is very difficult to know whether it's secure. It's very difficult to know whether the agents are acting in accordance to the boundaries that have been put in place, and the rights that they've been granted.”

After talking about logistics, auditing and other aspects of implementation, Dutta ended with a lightning round, asking panelists to throw out a few words on the agentic future in general.

“Are you going to measure the cost of the fuel, or more importantly, the cost of the whole trip?” Fegueiredo asked.

“Think of multimodality,” Sharaf said, “voice, vision, devices and robots.”

“Year one: messy, and year five: invisible,” Madhira said.

“It will be everywhere,” Seseri said, rounding out the panel. “Five years from now, I think we should start talking about AGI.”

This, to me, was a relevant look at what we’re going to be doing with AI agent swarms. The collaborative potential of these “little claws” is going to be a major issue in enterprise, and in life, going forward.