What is loop engineering?

Accuracy note: the people, dates, and figures in the origin story below are recorded as of June 2026 — verify before relying. Every external claim has a row in SOURCES.md.

Definition

Loop engineering is the practice of running an AI coding agent in a governed, verifiable loop until a clearly defined goal is met. Where prompt engineering shapes a single turn, loop engineering shapes the outer loop: what the agent does, how it checks its own work, when it stops, and what it costs.

A useful one-liner (Addy Osmani, as of June 2026 — verify before relying): loop engineering is "designing the loop the agent runs in, not just the prompt it runs once." See SOURCES.md.

Where the term came from

The idea predates the name. A short, labeled timeline (as of June 2026 — verify before relying; sources in SOURCES.md):

When Who What
Jul 2025 Geoffrey Huntley popularized the bare "Ralph" loop — re-run the same prompt in a while true until done
Jun 2026 Steinberger a viral thread on running agents in long unattended loops
Jun 2026 Boris Cherny framing loops inside coding-agent tooling
Jun 2026 Addy Osmani named and defined "loop engineering" as a discipline
Jun 2026 Greg Brockman described a governed "Ralph loop++"

The throughline: a one-shot prompt became a loop, and the loop grew governance (stop conditions, cost caps, verification) until it was worth treating as its own engineering surface.

Prompt → context → harness → loop

The four things you actually design, from innermost to outermost:

Lineage: prompt to context to harness to loop

Loop engineering is mostly about the outer two rings: the harness and the loop. The rest of this guide is how to design them.


Next: /goal vs /loop basics →