Team of Teams

Small startups inside the enterprise

Ryan Petersen describes his prescription for scaling without ossifying as getting people single-threaded on a problem or metric so they do not have to coordinate with everyone and drown in meetings. His worked example decomposes gross margin: buying freight, pricing it, and selling it are three distinct skills, so he splits them into a freight-buying team building relationships with ship and plane owners, an algorithmic pricing system, and a sales team that sells at the given price.¹ He ties this to what he calls a wrinkle in the law of large numbers: if a big company is really a collection of small teams, each small team can still grow fast, so the whole can keep growing rather than slowing as it gets large. Garry Tan frames it as "decentralized autonomy" with "rules of engagement."

The design has a stated precondition. Petersen concedes, citing Parker Conrad's "compound startups" idea, that autonomous teams must remain connected through a shared data model and APIs, or there is no reason to have one company rather than twenty-six separate ones. This is his answer to the central scaling problem of growing headcount without growing coordination cost, which is the pressure described in coases-theory-of-the-firm: a firm grows only until internal coordination cost exceeds the benefit, so mechanisms that lower that cost raise the size a company can reach.

Immune system versus nervous system

Petersen frames why the model is fast with a biological metaphor. He describes the squad model as an immune system with no centralized control that senses a problem and attacks it, rather than a nervous system where the edge waits for instructions from the brain.¹ He nearly called the units "cells," for the same split-and-grow biology, but a cell is also a prison cell. In his account the customer becomes the boss, so the edge does not need to be told what to do. He also treats the model as fragile: he attributes his return as CEO to the period when Dave Clark ran the company and Flexport moved away from the squad model toward task-based efficiency, and, in his telling, lost much of its customer-centric ethos, until volumes stopped growing and the board concluded the growth engine was gone, a degradation examined in quality-vs-efficiency-service-ops.

A parallel cultural formulation

Brad Jacobs describes a related structure he calls a superorganism: a management team whose collective output exceeds the sum of its parts, structured like a beehive or a colony.² Where Petersen's version is primarily architectural, Jacobs presents his as primarily cultural, created through rituals and psychological safety. His operational vehicle is the Monthly Operating Review, which he describes as roughly ten hours with about 25 people and no presentation theater, where the agenda is set by the group ranking submitted questions rather than imposed by the CEO, runs on what he calls dialectical debate, "we present different ways of looking at a situation that partly overlap, and partly conflict, and then we, in a very scientific way, figure out what's right," and closes with recognition rituals. He treats 25 as a deliberate ceiling, large enough for genuine dialectical debate and small enough that people do not peacock or suppress candor.

The intelligence-layer variant

Jack Dorsey, in his account of restructuring Block, goes further in one respect: rather than building efficient relay channels between autonomous teams, he describes eliminating the relay function by making every artifact queryable directly, so no handoff or team boundary requires an API to cross. Read together, these are three points on a progression from traditional hierarchy, to API-connected autonomous units that still have short relay chains, to an intelligence layer that replaces the relay entirely. All three treat the organization itself as a system to be redesigned rather than staffed.

Open questions the founders leave

Petersen concedes the line between healthy autonomy and fragmentation is unresolved and depends on the connecting data model. Jacobs's review model is CEO-intensive and relationship-dependent, leaving open whether it scales past the inner circle without creating a privileged tier. Dorsey's model shifts the coordination cost from human relay to model governance, which raises an unanswered question about who curates the shared world model.