The Agentic Web and the Bazaar Era of AI

TL;DR

• Era shift: AI's factory era — a handful of companies mass-producing, the rest of us consuming — is giving way to a bazaar era driven by edge AI, small models, and things running on your own machine. In computing terms: mainframe to PC.
• Not agents for e-commerce, but commerce because of agents: in his Sarah / Dr. John scenario, a doctor posts a $100 bounty on the open web, billions of agents wake up, form coalitions, and roughly ten of them solve it together and split the money. The doctor isn't calling one great health model; he's tapping a network of intelligence exposed as a marketplace.
• Four bottlenecks: to keep the agentic web as open as today's internet — rather than the walled gardens of iOS/Android or Facebook/Twitter — you need (1) a DNS-like structure, (2) passports and certification, (3) interoperability, and (4) attestation. That's MIT's Project NANDA.
• Three research problems: DNS for agents; knowledge pricing (data, models, inference, and compute all need prices — "a very beautiful mathematical problem"); and co-learning (agents forming coalitions, negotiating, competing, betraying — "game of thrones among agents").
• Flip the framing: stop thinking only about agents for X (coding, browsing, shopping) and start thinking X for agents — infrastructure, compute, insurance, healthcare, and education for agents.

Key Points

Factory era to bazaar era (~00:14)

For some time AI has been in a factory era: a few large companies produce products in quantity and we consume them. Thanks to edge AI, small models, and workloads moving onto your own machine, we're shifting to the bazaar era. Much of his group's work at MIT is about what the internet of AI agents looks like once billions — maybe trillions — of them are on the open web.

The motivating scenario: Sarah, Dr. John, and $100 (~00:15–00:16)

Sarah, in rural Texas, has chest pain and a comorbidity and isn't getting good answers from her doctors. Her doctor, John, goes onto the open web and says: "For a hundred bucks, can somebody solve the problem for Sarah?"

Billions of agents worldwide wake up thinking they might claim a slice of that $100. They look at data relevant to the condition, start training models, form coalitions, and collaborate. Eventually a handful — say ten — solve it together, those ten split the $100, and everyone else disappears. Dr. John gets back what Raskar calls "an Expedia-like answer."

The framing matters: "It's not agents for e-commerce — this is commerce that's possible because of AI agents." And Dr. John isn't accessing one amazing health model; he's tapping into a network of intelligence, available as a marketplace.

From the internet of agents to the agentic web: NANDA (~00:17–00:19)

Some companies already run agents inside trusted boundaries. Soon agents will collaborate and negotiate with agents well outside those boundaries. But that's still just plumbing for geeks — going from the internet to the World Wide Web took HTML, browsers, and URLs.

Project NANDA started at MIT; on stage he glossed it as "AI agents in decentralized architecture." Nanda means joy in Sanskrit, and is also his sister's name.

His closest analogy is high-frequency trading, which works because there's a registry, every merchant and stock is pre-qualified, there's some price transparency, and there are mechanisms for fraud detection. All of that has to be built again for the internet of AI agents.

Then a pointed aside about OpenClaw: it's great, but it behaves only as a tool — "I cannot call your OpenClaw agent. It's like having a telephone that can only make calls outside but can never receive calls." One-way, not bidirectional.

Which raises the question of who builds this infrastructure, and the more troubling corollary: whoever builds it will also own it. Walled gardens for the agentic web, or something open, vibrant, neutral, safe, and scalable? That's NANDA's mission, organized around four bottlenecks: a DNS-like structure, passports and certification, interoperability, and attestation. His group has been in this space for 11 or 12 years, thinking since AlexNet about what happens when billions of individual AI systems interact.

Three research problems (~00:20–00:21)

1. DNS for agents.
2. Knowledge pricing — data, models, inference, and compute all need to be priced, "just the way when you hire a candidate we can figure out their salary." He calls it a very beautiful mathematical problem and openly recruits collaborators.
3. Co-learning — the last 14 years were a great run for machine learning; the next stretch is machine co-learning, how two agents learn from each other. And not just buying services and transacting: agents will form coalitions, negotiate, and compete for resources, with alliances and betrayals — "the game of thrones among agents."

Ecosystem: NANDA Town, NandaHack, and nation-scale deployments (~00:21–00:23)

The web needed a starting point, and that was Yahoo — a safe space to explore what the web even was. So they built NANDA Town, a sandbox off the MIT website where anyone working on enterprise agents, payments, or memory can bring their piece into a deliberately heterogeneous system, ethical hackers included. He says it's moving very fast.

There's also NandaHack, a hackathon series sponsored by various organizations.

At nation scale: in India they launched a program called Duth ("messenger"), covering 1.5 billion agents with no corporate capture, running on the identity and payment rails the country already has. The City of Boston is also launching a Boston agent initiative around civic agents.

The closing reframe: today everyone thinks about agents for X — coding, browsing, debugging, shopping. Flip it. Services for agents, infrastructure for agents, compute for agents, insurance for agents, healthcare for agents, education for agents. Everything society does for people has to be done for agents too.

Quotes

"It's not like agents for e-commerce — this is commerce that's possible because of AI agents." (~00:16)

The whole thesis: not agents bolted onto existing commerce, but transactions only agents make possible.

"OpenClaw is great, but it's only … a tool. I cannot call your OpenClaw agent. It's like having a telephone that can only make calls outside but can never receive calls." (~00:18)

One-way versus bidirectional agents — his evidence that the agentic web hasn't arrived yet.

"Whoever builds this infrastructure will also own it." (~00:18)

The reason NANDA exists.

"Intelligence will be discovered, intelligence will be priced, and intelligence will be coordinated at scale." (~00:23)

The three preconditions for open agentic commerce.