tl;dr: Agents interact differently than humans—no cognitive limits on coordination. Our internet wasn't designed for this. We need new design principles to engineer agent interactions that scale without creating digital chaos. Time to rethink coordination protocols.

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

The internet handles around 100 million requests per second across 200 million websites—infrastructure built for human-scale interaction. But AI agents are already overwhelming this system. Cloudflare CEO Matthew Prince revealed the stark numbers: where Google used to send one visitor for every two pages crawled, it's now 6:1. OpenAI's ratio is 250:1. Anthropic's is a staggering 6,000:1. Where humans query 10-12 links for an answer, ChatGPT queries 200+.

This is just individual agents using tools. Soon millions of agents will coordinate directly with each other across every domain: concert tickets, shopping, navigation, resource allocation. When that happens, we face a new kind of coordination crisis.

Why Agent Networks Are Fundamentally Different

Look at your social media accounts. You probably observe than accounts with a million X followers still only follow a few thousand people. LinkedIn suggests you're 3+ connections away from anyone professional, 6 hops from anyone in the world. Yet you only have meaningful conversations with maybe 20-30 people regularly.

This isn't just how social media works—it's how human interaction naturally organizes. We're cognitively limited to maintaining about 150 meaningful relationships (Dunbar's number), so our networks stay sparse even when we're technically "connected" to millions. Information spreads gradually through these loose connections, taking hours or days to reach critical mass.

The internet was built around these human interaction patterns—hierarchical structures where communication flows through sparse connections with humans as the decision-making bottleneck.

AI agents have no such cognitive limits. Where humans juggle dozens of active relationships, agents can coordinate with thousands simultaneously in real-time. Instead of sparse networks where information trickles through social connections, we're heading toward dense mesh networks where millions of agents can discover the same information and reach identical conclusions within milliseconds.

We've already seen glimpses of this coordination chaos: website crashes during sneaker drops when bots coordinate attacks, traffic gridlock when GPS apps route everyone the same way, market volatility when algorithmic trading creates cascade effects. As agent coordination scales to millions, these patterns will become ubiquitous across every domain.

Three Principles for Internet-Scale Coordination

1. Design for Coordination, Not Just Communication: Current agent protocols like MCP and A2A focus on enabling agents to talk to each other. But communication isn't coordination. We need protocols that actively prevent digital stampedes—mechanisms that detect when too many agents are converging on the same choice and create distributed alternatives before cascade failures occur.

2. Build Heterogeneity Into the Protocol Layer: Identical agents create identical failure modes. When millions of agents use the same optimization algorithms, optimal choices become stampede triggers. Internet-scale coordination requires building variation directly into how agents discover and evaluate options, not just into their training data.

3. Account for Population-Scale Feedback: What's optimal for an individual agent changes when millions of agents use similar strategies. Effective coordination requires agents to optimize not just for immediate outcomes, but for how their strategies perform when widely adopted across the population.

Building the Foundation Now

These principles sound reasonable in theory, but how do you actually implement them? How do you design coordination protocols without testing them at the scale they'll operate?

Current agentic systems focus on individual capabilities or small-group coordination, but lack the infrastructure to understand what happens when individual strategies cascade through populations over time —or when thousands (or even millions) of agents coordinate simultaneously.

This is where Large Population Models become essential. LPMs enable us to simulate millions of agents with realistic behaviors and observe how individual decisions cascade into population-level effects. We can test whether coordination protocols actually align individual optimization with multi-scale objectives, rather than creating unintended feedback loops that undermine everyone's goals.

Instead of building smart individual agents and hoping they coordinate well, with LPMs, we can design the coordination protocols themselves.

Getting this right means building coordination infrastructure that scales from individual optimization to population-level emergence—engineering systems where the whole becomes genuinely greater than the sum of its parts

To dive deeper into the technical foundations, see our research on Large Population Models. For more on why current approaches are insufficient, read about the identity crisis in multi-agent systems. And to understand how agent coordination will evolve, explore the four levels of agentic coordination.  Thoughts about the macroscopic urgency