The Geometry of Networks

Not All Networks Scale the Same Way

In the previous post we introduced coordination density — the ratio of productive interactions to the cost of maintaining them — and showed how Metcalfe’s Law, Brooks’s Law, and Conway’s Law shape it inside the modern firm.

But that analysis assumed a single type of network. It treated all connections as equivalent.

They are not.

A television audience is a network. So is a telephone system. So is an open-source community. But they scale in fundamentally different ways, because the geometry of their connections differs.

The geometry of a network determines how its value grows — and which institutions it produces.

Network theory gives us four useful models, each with a distinct scaling law.

Four Network Geometries

1 — Hierarchical Networks (The Firm)

Traditional firms are trees. Decision-making and information flow vertically. Cross-node communication is deliberately limited to reduce coordination cost.

The tree is a constrained network. It sacrifices connectivity for control. Each node connects only to its parent and children, so the number of active connections grows linearly with the number of nodes.

Connections ∝ N

This is the topology Brooks’s Law warned about: adding nodes to a tree adds management overhead without adding cross-cutting connections. The hierarchy scales reach but not interaction.

2 — Broadcast Networks (Sarnoff)

Broadcast networks have a central sender and many receivers. Information flows one way.

David Sarnoff observed that the value of these networks grows roughly in proportion to the size of the audience.

Value ∝ N

Television, radio, newspapers, streaming platforms. Double the audience, roughly double the value. Broadcast networks scale, but slowly.

3 — Connection Networks (Metcalfe)

Connection networks allow participants to communicate with each other. Every new participant creates potential connections with every existing one.

Telephone networks, email, messaging systems, social networks. Robert Metcalfe observed that the value of these networks grows roughly with the square of the number of users.

Value ∝ N²

This is why communication networks become extremely valuable once they reach critical mass — and why firms that become connection networks unlock value that hierarchies cannot.

4 — Group Networks (Reed)

Some networks go further. They allow participants to form groups — ad-hoc subsets that collaborate, dissolve, and recombine.

Open-source communities, Slack workspaces, professional communities, research collaborations. David Reed observed that group-forming networks scale even faster because the number of possible subgroups grows exponentially.

Value ∝ 2ᴺ

A network of 20 participants can theoretically form over one million possible groups.

But Reed’s Law describes a theoretical ceiling, not a guarantee. Most possible subgroups are trivially small or valueless. Real group-forming networks face coordination costs within groups — scheduling, context-sharing, trust-building. What matters is not the raw number of possible groups but the activation rate: the fraction of potential groups that actually form and produce value. This makes the mechanisms for group formation — tooling, norms, protocols — as important as the network’s size.

How They Compare

Small differences in network geometry produce enormous differences in value at scale.

The Geometry of Institutions

These geometries map to different institutional structures.

Institutions differ not only in incentives and governance — but in network geometry. And that geometry determines how they scale.

This explains something that pure Coasean analysis misses. When we ask “why do firms exist?”, the answer is not only about transaction costs. It is also about which network geometry best fits the coordination problem at hand.

Hierarchies dominated when coordination was expensive and bandwidth was scarce. They deliberately constrained the network to make coordination manageable. But as coordination costs fall, the constraint becomes the bottleneck — and institutions shift toward geometries that unlock more connections.

A Case Study: The Legal Industry

The legal industry illustrates this shift concretely.

The Hierarchical Firm

Historically, legal work required scarce expertise, extensive research, and high coordination costs. The traditional law firm organised this as a tree.

Work flowed up and down the hierarchy. Clients hired the firm, not individual lawyers. The tree managed coordination costs by limiting who could interact with whom.

The Connection Platform

Digital platforms are restructuring this as a connection network. Legal marketplaces match clients directly with lawyers across the network.

Value scales faster because the number of possible matches grows quadratically. The firm’s role as intermediary weakens.

The Group-Forming Network

The most interesting development is the emergence of dynamic specialist teams — groups that assemble around a specific matter, draw on shared research and AI-assisted analysis, then dissolve and recombine for the next problem.

This is a group-forming network. The same participants form different subgroups depending on the task. Each subgroup is a temporary coordination unit that produces value no single node could produce alone.

The geometry has changed — and with it, the boundary of the firm.

Agent Networks: A Geometric Shift

Until recently, all four geometries were constrained by a single bottleneck: human bandwidth.

Humans can maintain a limited number of active connections. They fatigue, forget context, and need synchronisation rituals — meetings, standups, status reports. These constraints cap the practical density of any network, regardless of its theoretical geometry.

AI agents remove this constraint.

Agents do not fatigue. They can maintain thousands of concurrent connections. They can join a group, contribute, and move on without context-switching costs.

This changes the geometry of the network itself.

Agents do not just add N. They change the type of scaling.

Consider the Cyborg Cell from Post 9 — one human anchor plus N specialist agents. Within the cell, every node connects to every other node. The cell is a fully connected mesh. That is already Metcalfe-scale coordination operating at the level of a single worker.

Now consider multiple cells interacting. Each cell can form working groups with other cells — an IP specialist cell collaborating with a regulatory cell and a research cell for a specific matter, then dissolving and recombining for the next one.

Within cells: Metcalfe dynamics (fully connected mesh). Between cells: Reed dynamics (group formation and dissolution).

The agent network does not merely add more nodes to the existing geometry. It shifts the firm from a constrained hierarchy toward a group-forming network — potentially unlocking Reed-scale value.

The Coordination Problem

But this creates an immediate problem.

Recall from Post 10 that coordination density measures the ratio of productive interactions to coordination cost. Reed-scale networks have enormous potential — but they also have enormous coordination burden. A million possible groups means a million possible coordination failures.

Hierarchies solved this problem by limiting connections. A manager decided who worked with whom.

Group-forming networks cannot rely on that mechanism. They need something else — a way to coordinate interactions without a central authority deciding each one.

Dense networks need coordination mechanisms that are not hierarchy.

The question is no longer how many connections can we create but how those connections are governed.

From Geometry to Governance

The history of institutions can be read as a sequence of network geometries, each unlocking more connections — and each requiring new coordination mechanisms.

Each transition reduced coordination costs and increased network density. Each required a new mechanism to prevent the network from collapsing under its own complexity.

The agent network is the latest transition. It unlocks Reed-scale group formation. But its coordination mechanism is not yet established.

Network geometry determines scaling potential. But potential without coordination collapses into noise.

That coordination mechanism — sitting between the participants and the infrastructure, governing how interactions happen without requiring hierarchy to manage each one — is the protocol layer.

References & Intellectual Lineage

• Sarnoff, D. — broadcast network value scaling
• Metcalfe, R. (1980). Network value and telecommunications economics.
• Reed, D. (1999). “The Law of the Pack” — group-forming network scaling.
• Brooks, F. (1975). The Mythical Man-Month.
• Conway, M. (1968). “How Do Committees Invent?”
• Coase, R. (1937). “The Nature of the Firm.”
• Post 9 in this series: The Hybrid Topology — the Cyborg Cell.
• Post 10 in this series: Metcalfe’s Law, Conway’s Law, and the Networked Firm — coordination density.