Pattern Engineering

Shaping the loops, closing the loops

Hayden White on X3 min read

I began to notice a pattern of patterns in today’s age, which has had me thinking ever since.

A framed organization chart of the Tabulating Machine Company, dated December 1917.
The Tabulating Machine Company’s organization chart, December 1917. Photo by Marcin Wichary, CC BY 2.0.

As the volume of agent work explodes, engineering leverage will concentrate in two areas. One is already getting attention, the other is less clearly named.

⟣ Verification engineering ⟢ (closing the loop): building the tools, systems, and processes that can reliably establish whether a change meets its intended requirements, and produce actionable evidence and signals when it doesn’t.

⟣ Pattern engineering ⟢ (shaping the loops): designing, curating, and reinforcing the conventions, abstractions, tools, and context that improve subsequent work across a codebase and company.

These practices have always mattered, but agents change how frequently those decisions get applied. They repeatedly learn from and extend the patterns already present in a codebase. A good pattern can improve thousands of subsequent changes, while a bad one can repeatedly steer work in the wrong direction.

Henry Binney Hawke at his workbench, finishing a wooden pattern for a gear wheel.
Henry Binney Hawke finishing a wooden pattern for a gear wheel, to be cast in his foundry in Kapunda, South Australia, circa 1870s. State Library of South Australia.

What it looks like in practice

Pattern engineering means removing sources of recurring misdirection: unit tests that lock in implementation details, stale architecture specs, unclear service ownership. It also means making good work easier: enforcing clear architectural boundaries, making good observability and tracing automatic through shared abstractions, and establishing which context is worth preserving, where it belongs, and when it needs updating.

My favorite example

One pattern that has served our engineering team very well: a single reliable way for a human or cloud agent to configure secrets, run your stack, and test a change in any supported development environment.

A work trial candidate on a fresh Mac and a cloud agent should follow the same setup workflow, with no undocumented steps. Once dependencies and access are provisioned, you should have a running stack in under a minute.

And that workflow should stay current as engineers and their agents add services. It should derive from the same service definitions they already maintain, so supporting a new service doesn’t require someone to remember to update a separate setup system. Every new engineer and agent benefits from that work, and every new service extends it.

revup, the CLI we use to set up every engineer's and agent's machine at Revyl

Pattern engineers

I think of people like @poteto and @dexhorthy as pattern engineers. Their work turns engineering judgment into principles and workflows that other people and agents can reuse. pstack turns engineering patterns, including verification itself, into reusable agent skills. That’s what makes it so useful. It’s a crash course in pattern engineering with the batteries included.

Two companies, side by side

Now put two companies with access to an unlimited array of agents side by side.

One repeatedly spends time trying to rediscover setup, correcting the same mistakes, figuring out whether work is done. The other turns those lessons into reliable tools, defaults, and verification pipelines.

Car chassis moving down the assembly line at Ford's Highland Park plant.
The assembly line at the Ford Motor Company’s Highland Park plant, 1913. Library of Congress.

As the volume of agent work grows, that difference compounds massively. The best teams know where to invest.

Hayden White

Published November 18, 2026