Billions of AI agents will reshape society—and why we need to start measuring before it happens

- Ayush Chopra, PhD Candidate MIT (media.mit.edu/~ayushc)

Every ten years, America conducts one of democracy's most important rituals: the Census. We count every person, map every household, and catalog how 330 million Americans live, work, and move. This data becomes the foundation for $2.8 trillion in federal spending annually through 353 federal programs—from schools and hospitals to roads and disaster relief.  

But we're approaching a fundamental shift that will dwarf the complexity of counting 330 million Americans. Within the next decade, we'll need to coordinate not just billions of humans, but tens of billions of AI agents acting on their behalf.  Welcome to the era of the Agentic Census.

The Agent Revolution Coming

Right now, AI agents are still mostly personal assistants. Software engineers use GitHub Copilot to write code. Families ask ChatGPT for homework help. People use Claude for creative writing or to analyze their symptoms before doctor visits. These feel like individual productivity tools—helpful, but not transformative.

But look closer. Every month, these agents get more capable. They're starting to handle more complex tasks, coordinate between different systems, and take actions on our behalf. What begins as "AI helps me write emails" becomes "AI manages my entire communication workflow." What starts as "AI helps debug my code" becomes "AI handles most of our software development pipeline."

We're in the early stages of the biggest coordination revolution in human history—and we have no infrastructure to measure it.

The Cascade Effect Coming

Here's what keeps me up at night: When AI agents begin coordinating large parts of the economy, they won't just affect individual jobs—they'll create cascading effects throughout society that no one can predict.

Let me paint a picture of what could happen in North Carolina over the next few years.

AI agents start handling routine software engineering tasks in the Research Triangle. At first, this just makes programmers more productive. But as the agents get better, companies need fewer engineers. Some developers lose jobs, others relocate to where the remaining high-level work is. Local restaurants and services that depended on tech worker spending see revenues drop. Housing demand shifts. The service workers who supported the tech ecosystem—from baristas to daycare providers—face their own disruptions.

Meanwhile, some new jobs emerge: AI coordinators, human-AI collaboration specialists, complex problem solvers who work alongside advanced agents. But these jobs might be in different places, requiring different skills, serving different communities.

The entire economic ecosystem of the Research Triangle could reorganize around AI coordination in ways no one can predict—because we have no way to see it coming.

Why Traditional Measurement Fails

Our current statistical systems are built for the economy we used to have. The Bureau of Labor Statistics tracks human employment but has no framework for measuring AI agent capabilities or their economic impact. The Census Bureau maps where people live but won't track how agents coordinate where people work, learn, and receive services.

But there's a deeper problem: Traditional statistics measure outcomes after they happen. They're like trying to understand a car crash by examining the wreckage—you can see what broke, but you can't predict the next accident.

What we need is the ability to simulate how humans and agents will interact before those interactions reshape entire communities.

The Missing Layer: Agent Population Data

At MIT, we've built Large Population Models (LPMs) that can simulate millions of individual humans interacting in realistic economic environments. LPMs already work with traditional census data—they can model every software engineer in the Research Triangle, every restaurant that serves them, every service worker in the ecosystem. They can run thousands of scenarios to see how economic changes might cascade through regional communities. [What are LPMs?]

But there's a critical missing piece: While LPMs can simulate human populations using existing census data, they can't simulate human-agent interactions because we don't have census data for agents.

Think about it: The traditional census tells us there are 50,000 software engineers in North Carolina. But how many AI coding agents are being deployed? What can they do? How are they coordinating with human workers? We're trying to simulate a hybrid human-agent economy with data about only half the population.

Building the Missing Infrastructure

This is why we're building NANDA Registry—to index the agent population data that LPMs need for accurate simulation. Just as traditional census works because people have addresses, we need a way to track AI agents as they proliferate.

NANDA Registry creates the infrastructure to identify agents, catalog their capabilities, and monitor how they coordinate with humans and other agents. This gives us real-time data about the agent population—essentially creating the "AI agent census" layer that's missing from our economic intelligence.

Here's how it works together:

Traditional Census Data: 171 million human workers across 32,000+ skills
NANDA Registry: Growing population of AI agents with tracked capabilities
Large Population Models: Simulate how these populations interact and create cascading effects

The result: For the first time, we can simulate the full hybrid human-agent economy and see transformations before they happen.

The Agentic Census in Action

We've already built a working demonstration. Project Iceberg combines traditional census data on human workers with emerging agent capability data from NANDA Registry, then uses LPMs to simulate workforce transformation across all 50 states.

The results reveal exactly the kind of intelligence an Agentic Census provides: We can see that North Carolina faces potential automation pressure affecting $12.8 billion in economic activity across 116,000 workers. But more importantly, we can simulate the cascading effects—which communities will be affected, what new opportunities will emerge, where re-skilling programs should be targeted.

This is the Agentic Census working: Real-time data about both human and agent populations, combined with simulation to reveal how they'll interact and reshape society. 

Check out our fireside with CNBC's Mackenzie Sigalos and Iceberg launch blog.

The Sovereignty Question

Here's what worries me most: If we don't build public infrastructure for the Agentic Census, private platforms will control that intelligence exclusively.

The companies building AI agents will also control the data about how those agents reshape society. They'll see economic disruptions coming before local leaders do. They'll understand workforce transitions before the workers experiencing them. They'll run simulations we can't access using data we don't have.

Without public Agentic Census infrastructure, we're potentially headed toward a future where the most important decisions about our communities are made by algorithms we can't see, based on simulations we can't run, using agent population data we don't control.

The Window Is Closing

We're in a unique moment. AI agents haven't yet proliferated enough to reshape major economic systems, but they're advancing rapidly. We have perhaps 2-3 years to build the agent population tracking infrastructure before the transformation accelerates beyond our ability to measure it.

The technology exists today. At MIT, we've proven that LPMs can simulate human-agent interactions when given adequate data. NANDA Registry is operational and beginning to track the emerging agent population. Project Iceberg demonstrates how this infrastructure works for workforce planning.

The question is whether we'll deploy this infrastructure before we desperately need it.

The states that build Agentic Census infrastructure now will have economic intelligence ready when AI agents begin coordinating major parts of their economies. They'll see transformations coming and guide them proactively. The states that wait will find themselves managing changes they can't understand, with data systems designed for an economy that no longer exists.

The invisible economy isn't here yet. But it's coming fast. The question is whether we'll build the infrastructure to see it—and simulate it—before it reshapes everything we think we know about how society works.

What are Large Population Models? : LPM overview

Learn more about the Agentic Census at iceberg.mit.edu and join.projectnanda.org