Notes from the NANDA Summit at MIT Media Lab: why intelligence lives in the interaction between agents, and why leadership has no visibility into what agents already run inside their companies.

I spent yesterday in Cambridge at the NANDA (nanda.mit.edu) Summit, hosted by Prof. Ramesh Raskar's MIT Media Lab group with HCLTech. It's been great watching NANDA grow into one of the sharpest gatherings on where agentic AI is actually headed, since it's inception in early 2025.

One moment from the day reframed how I think about the whole category.

Ayush Chopra, a PhD researcher at MIT Media Lab, put up two examples side by side. A group of humans trying to maneuver an oddly shaped object through a narrow gap. A colony of ants doing the same task. Individually, the humans are smarter by every measure we have. But add more humans to the group and performance gets worse. Add more ants and performance gets better.

His point: intelligence doesn't live in the individual agent. It lives in the interaction between agents.

He backed it up with data. In a resource-sharing simulation, a swarm of small, cheap AI models, coordinated properly, beat a single large model working alone, at a fraction of the cost. The same approach is now in production: a live supply-chain-risk system gave 180 days of early warning before a disruption showed up in conventional data, and a related MIT project, Iceberg, tracks over a trillion dollars of U.S. labor market risk in real time.

If that holds up outside a research paper, it changes the default architecture question for anyone building with agents: not "which is the best model," but "how do many smaller agents need to talk to each other."

Two other things stayed with me.

Every enterprise panel circled back to the same blocker, and it isn't the technology. Leadership has no visibility into what agents are already running inside their companies, what those agents can touch, or what they're doing with that access. Multiple panels arrived at the same fix independently: a registry every agent must be enrolled in, and machine-readable policy agents can be checked against automatically, since no human can review every decision at scale.

And a reassuring data point, from a manipulation experiment. Researchers tried to bias an AI purchasing agent by injecting fake negative reviews about one vendor and glowing ones about a competitor. The agent ignored it and stuck to the facts, a real trust advantage over human buyers, who are demonstrably swayed by that kind of framing.

One slide put a number on the stakes: agent-to-human ratio, one million to one. That's the world the organizers are planning for, from city governments navigating permits, to agent access for 20 million people at India's Kumbh Mela, to a future grid with 200 million generators trading energy automatically. Same question every time: once agents act on our behalf at that ratio, who's watching them, and how do we know they're still working for us?

That question will define this technology's next phase more than any model release.

Curious what others in this space are seeing on the ground.