Who Will Own the AI Agent Economy? MIT's Warning to Builders

MIT's Ramesh Raskar says the window to keep the agent web open is closing. Here is what the shift from "agents for X" to "X for agents" means for founders, and why the next decade of AI depends on who owns the rails.

The mainframe moment for AI agents

MIT Media Lab professor Ramesh Raskar has a line that should be pinned above every AI product team's desk: "Right now, we are in the mainframe era of AI. We are moving to the PC era of AI, and that changes everything."

In an MIT Sloan talk from the AI + X series, Raskar — who leads Project NANDA, MIT's research initiative on the "internet of AI agents" — laid out a near-future where every person, organization, fridge, car, stock, and city has its own AI agent. Billions, then trillions, of agents transacting, negotiating, and coordinating on our behalf.

The question he asked the room is the one every builder should be asking this year: who is going to own the rails those agents run on?

From "agents for X" to "X for agents"

Most of what ships today is what Raskar calls "agents for X" — an agent for stock trading, an agent for customer support, an agent for booking travel. Useful, but small.

His inversion:

> "The economy of the future is not an economy where we say, 'Let's create agents for X.' It's an economy where we instead say, 'Let's create X for agents.'"

That flip changes the opportunity set completely. If trillions of agents are about to act on the internet, the biggest businesses will not be the agents themselves — they will be the marketplaces, protocols, and services those agents need to function.

Think of it the way AWS ate the "build a website" era. The gold was not in the websites. It was in the compute, storage, DNS, CDN, and payments layer underneath.

Four categories where the agent economy will be built

Raskar highlighted four infrastructure categories that don't really exist yet — and that will need to exist for an open agent web to work.

1. Agent identity and discovery

An ICANN for agents. Right now there is no universal registry that answers "who is this agent, who owns it, what is it allowed to do, and how do I reach it?" Without that, agents can't safely find or trust each other across company boundaries. Project NANDA is one attempt; expect standards fights here for years.

2. Trust and reputation

Agents will impersonate other agents. They will lie. They will get hijacked. The web needs the equivalent of passports and credit scores for agents — verifiable identity, reputation histories, and revocation lists. Whoever owns this layer becomes the Visa of the agent internet.

3. Insurance, repair, and legal services

Agents will make mistakes at scale. They will misprice trades, miss deadlines, and get sued. Raskar predicts a whole services layer will emerge: agent insurance, agent maintenance, agent arbitration. Boring on the surface, enormous underneath.

4. Stablecoin micropayments

Human payment rails were not designed for a trillion tiny machine-to-machine transactions per second. Raskar sees stablecoin-based micropayments as the natural settlement layer — instant, programmable, cheap enough that an agent can pay 0.003 cents to another agent for a database query without blinking.

This is the same thesis Stripe, Circle, Visa, and Coinbase are quietly aligning behind. Our recent post on Stripe's agentic payments infrastructure walks through what that looks like in practice.

Open web vs. walled garden — which future are we heading to?

Raskar frames the choice through history.

Scenario A — the open web repeat. The World Wide Web became universal because its foundational layer (HTTP, DNS, TLS, HTML) was interoperable and beyond any one company's control. Anyone could build. Anyone could plug in. That produced Google, Amazon, Wikipedia, Shopify, and millions of small businesses.

Scenario B — the smartphone repeat. iOS and Android don't talk. Keynote and PowerPoint don't talk. The two-app-store duopoly decides what ships. Innovation happens inside the walls those companies drew. If the agent economy evolves this way, you'll get whatever agents Apple, Google, Microsoft, Meta, OpenAI, and Anthropic decide to ship you. Nothing else.

Raskar is candid: "Nine out of ten paths we take will lead us toward AI agents consolidated under corporate control." His work at NANDA is a deliberate push against that gravity — because "the window to keep this web of agents open is closing soon. But it's not closed yet."

What this means for founders and operators right now

If you are building in AI, the MIT thesis has three immediate implications.

1. Stop competing agent-for-agent with the frontier labs. OpenAI, Anthropic, Google, and Microsoft will ship general-purpose agents that are better than anything a small team can build. Do not race them on their turf.

2. Build the picks and shovels. The durable, defensible businesses of the next five years are the ones supplying the agent economy: identity, verification, payment rails, tool marketplaces, evaluation, observability, sandboxing, insurance, dispute resolution, and vertical protocols. Boring, technical, sticky.

3. Design for interoperability first. If you ship an agent product, assume it will need to talk to other agents built by other companies. Support MCP (Model Context Protocol), open agent-to-agent handshakes, and portable identity from day one. The teams that lock users into a proprietary stack will look right for 18 months and dead by month 36 — the same story that played out with proprietary APIs in the Web 2.0 era.

The 88 Labs read

MIT's framing lines up with what we see building agents for real businesses every week:

The frontier is not the model — it's the middleware. Deployments succeed or fail on identity, permissions, payments, memory, and observability. That is exactly the layer Raskar says is missing.

Vertical wins the short game. Protocols win the long game. A great real-estate voice agent will make money next quarter. A protocol that lets any real-estate agent talk to any MLS agent will define the industry.

Open beats closed when the standards are set early enough. The next twelve to twenty-four months is when the agent web's HTTP moment happens. Whoever ships an open, credible standard now writes the rules.

FAQ

What is the "AI agent economy"?

The AI agent economy is the emerging market of software agents that transact, coordinate, and act on behalf of people, companies, and devices — plus the infrastructure (identity, payments, trust, insurance) those agents need to interact at scale.

What is Project NANDA?

Project NANDA is an MIT Media Lab research initiative led by Ramesh Raskar, focused on the protocols an "internet of AI agents" needs: identity, discovery, addressability, registries, verification, interoperability, and decentralized coordination.

Why does Raskar say "X for agents" instead of "agents for X"?

Because most of the economic upside will not be in individual agents, but in the marketplaces, protocols, and services that trillions of agents need to function — the way AWS captured more value than most websites did.

Will one company own the AI agent economy?

Raskar estimates nine out of ten paths lead to consolidation under a handful of large corporations, similar to the mobile OS duopoly. Open standards work like Project NANDA, MCP, and open agent-to-agent protocols are the main counterweight.

Where do stablecoins fit in?

They act as the settlement layer for machine-to-machine micropayments — programmable, near-instant, and cheap enough for trillions of small transactions between agents.

Bottom line

The agent economy is not a feature race. It is an infrastructure land grab, and MIT is right that the window to keep it open is closing fast. If you're building in AI in 2026, the highest-leverage question is not "what agent should I build?" — it's "what does every other agent need, and can I ship it before the walled gardens do?"