← Glossary

Loop Engineering

Building out the supporting infrastructure around a bare agent loop -- scheduled automations, worktrees, skills, plugins, subagents, and memory -- so that a simple 'run until the goal is reached' loop becomes a dependable, self-correcting coding harness.

Context

Loop engineering was named by Addy Osmani in his essay “Loop Engineering” and covered on Episode 30 of the ADI Pod. It starts from the Ralph loop — a /loop that runs an agent over and over until a goal is reached — and names the pieces that turn that bare loop into a real harness:

By this definition, both the Codex app and Claude Code are already full loop-engineering harnesses; the term is more generic than any one implementation.

Why It Matters

Shimin’s framing on the show traces a progression in how developers direct agents: prompt engineering (tell the agent what to do) → spec-driven development (define the end goal precisely) → loop engineering (keep the goal looser, but bake your judgment into reusable skills and tools). Because the loop turns out so much more code, it amplifies whatever judgment you implant: the metaphor on the episode ran handsaw → table saw (“watch your fingers”) → tree-harvesting heavy machinery, where a mistake risks far more.

The drawbacks Osmani names are the ones the show keeps returning to: verification still falls on the developer (skip it and the output drifts), and cognitive debt bites hardest when a self-running loop breaks at 3am and no one understands the repo anymore.