Dunbar's number and engineering orgs
Anthropologists and sociologists argue about every neat constant that reaches management books. Dunbar’s number has earned skepticism for the same reason it keeps spreading. It maps a real empirical pattern in primate neocortex size versus mean social group size onto humans by interpolation, then checks whether human institutions land in the same ballpark. The fit is suggestive, not a license to reorganize by headcount alone.
Humans are said to be able to attach names to around 2,000 faces but have a cognitive group size of only about 150.
That line is from Robin Dunbar’s 1998 piece on the social brain hypothesis. “Cognitive group size” is the number of individuals you can maintain as a coherent social circle with stable mutual knowledge, not the number of people who can appear on an all-hands slide. Later work asks whether online networks blow past the limit in practice; the short answer in one careful empirical probe is that the old band still shows up in where people put attention, even when the UI invites infinite friending.
Org design
Treat ~150 as a warning about coordination cost, not a target team size for a service mesh. It pairs with Conway’s law when the boundaries on a diagram do not match who actually talks. Above the band where everyone can plausibly know everyone else’s work context, you rely on structure (chapters, pillars, domains) and repeated rituals (reviews, incident reviews, tech talks) to substitute for ambient knowledge. When those rituals thin out, you do not get a linear slowdown. You get duplicated work, surprise ownership gaps, and politics that scale faster than output because people are guessing motives instead of shipping from shared facts.
If you are splitting a growing org, the useful question is not “are we past 150 yet” but “where did informal repair stop working.” Expect that fracture well before payroll hits three digits for a single product surface.
On-call
Pager load is only partly a scheduling problem. The rest is trust under uncertainty. A responder who knows who owns which subsystem, who has shipped risky changes lately, and who will back them on a bad call makes better page decisions at 3 a.m. than a perfectly fair rotation across strangers. Dunbar-shaped groups (stable crews, long-lived rotations, explicit shadowing) trade statistical fairness for lower variance in outcomes. Larger pools can still work if you invest in runbooks, observability, and handoff discipline so the cognitive work moves from memory into systems. If you skip that investment and only widen the rotation, you are simulating a bigger group without adding bandwidth.
Communication overhead
Every working relationship is an edge. Possible pairwise edges grow with the square of headcount while calendar time does not. Layering exists partly so most edges are short (you talk mostly to your squad) and long edges are mediated (leads, RFCs, internal newsletters that compress decisions into narratives). At scale, managers are lossy compression for the graph.
None of this makes 150 a magic ceiling. It does explain why “just add another ten people” stops feeling linear, and why the fixes are structural instead of motivational posters.