Evolution with Scale

Overview

Architecture is not one job done at different volumes. The work changes in kind as an organisation grows, and the changes arrive at fairly predictable points.

Knowing where those points are is what lets you tell the difference between a problem you should solve now and one that belongs to a company you are not yet.

Three regimes

Up to roughly fifty engineers. Architecture is a part-time activity of the strongest engineers, and that is appropriate. Everyone can hold the whole system in their head, decisions propagate by conversation, and a dedicated architect would mostly add latency. What breaks first is not the architecture but the shared understanding — the point where no single conversation reaches everyone who needs to know.

Around a hundred to two hundred. The regime where the role appears, and Module 8 established why: nobody holds the whole picture, so cross-cutting decisions stop happening by themselves. Work that was previously invisible — recording decisions, maintaining a description of the system, sequencing structural change — becomes somebody’s job because it has stopped being nobody’s.

Five hundred and beyond. Architecture becomes a function rather than a person. The problems shift again: not making good decisions, but making consistent ones across groups that never speak, and building the platform layer that stops every team solving the same problem differently. This is the regime that boards, standards and radars were designed for, which is why Module 9 deferred them.

graph LR
  A["<b>~50</b><br/>shared understanding<br/>breaks first"] --> B["<b>~150</b><br/>nobody holds the<br/>whole picture"]
  B --> C["<b>500+</b><br/>consistency across<br/>groups that never meet"]

  classDef sm fill:#7D9A4E,stroke:#5F7A38,stroke-width:2px,color:#ffffff
  classDef md fill:#E8A33D,stroke:#C4832A,stroke-width:2px,color:#ffffff
  classDef lg fill:#5B8FB9,stroke:#4A7A9E,stroke-width:2px,color:#ffffff
  class A sm
  class B md
  class C lg

Three regimes, and what fails at each boundary. The failure is never capability — it is always a communication structure that has stopped reaching far enough.

Why the role appears when it does

Not because systems get complicated — a fifty-person company can have a genuinely complex system. It appears because the cost of an uncoordinated decision exceeds the cost of coordinating it, and that crossover is a function of how many people can make decisions independently.

Which also tells you what the role is for. Not being the best designer in the building; being the thing that makes a decision reach the people it constrains.

Anticipating without building prematurely

The dangerous half of this topic. Knowing what breaks at the next transition invites building for it now, and that is how organisations acquire platform teams for problems they do not have, abstraction layers over single implementations, and governance that fits a company four times their size.

Module 8’s test is the one to apply: what is cheap now and expensive to retrofit? Very little qualifies, and the things that do are mostly conventions rather than code —

Boundaries that follow the domain rather than the order work happened. Cheap to get right while a service is young; the single most expensive thing to change later.
Decision records. Free to start, impossible to reconstruct afterwards, as Module 9’s recovery topic established.
Observability and cost attribution per unit of work. Cheap while there are few services; a retrofitting project once there are many.

Everything else — the platform layer, the standards catalogue, the review board — is cheaper to build later with real requirements than now with imagined ones. The honest default is to let the next stage’s problems arrive and be solved by whoever is there, which is the opposite of what an ambitious architect wants to hear.

One hundred and forty, and what the number implies

Around 320 employees, about 140 of them engineers, in roughly eighteen teams. Founded in 2016; about $48M in annual recurring revenue.

So the company sits exactly at the transition where the role appears. That is why the job exists, and it explains features of the situation that otherwise look like dysfunction. There is no architecture practice because until recently none was needed. Decisions propagate by talking to Ivan Petrov because for years that genuinely worked. The twenty Go services have boundaries that follow extraction order because at fifty engineers nobody needed to coordinate that — and Module 8’s Conway analysis found exactly this.

None of that is incompetence. It is a company whose coordination structure fit it until roughly a year ago.

What the number predicts about what breaks next. Eighteen teams is past the point where one person can hold the system, and the symptom is already visible: nobody can say why the service boundaries are where they are, and reconstructing it would take weeks of one principal engineer’s time. That is the ~150 transition arriving on schedule.

And what it says about the AI platform. The initiative adds two more services to an estate whose boundaries are already undocumented, during a migration that consumes all capacity. The scale lens says the right investment is not a platform layer or a governance board — it is the cheap-now-expensive-later list: domain-shaped boundaries for the two new services, decision records for the choices being made this quarter, and per-interaction cost attribution built in from the start rather than retrofitted when the CFO asks.

What it says to defer. Anything sized for 500 engineers. The AI platform does not need a self-service developer platform, a service mesh, or an internal model gateway serving many teams. One team, two services, no spare capacity: the next stage’s problems belong to the next stage.

Feeds the workbook

The scale reading shapes your migration sequence and your legacy strategy — specifically, which structural investments you make now because they are cheap now, and which you explicitly defer to a larger company.

→ The regime is clear. The useful question is what gives way first. — What breaks at the next transition