Harness engineering

Why read it

This is the essay that named the discipline. Ryan Lopopolo (OpenAI) describes how a team of three engineers shipped a product with 0 lines of manually-written code — every line written by Codex — reaching ~1 million lines across ~1,500 PRs in five months, at ~3.5 PRs/engineer/day. Throughput went up as the team grew to seven.

The core claim: humans steer, agents execute. The engineer's job is no longer to write code but to design environments, specify intent, and build feedback loops that let agents do reliable work.

What's inside

• Repo as system of record — AGENTS.md is a table of contents (~100 lines), not an encyclopedia. Knowledge lives in a structured docs/ tree with design docs, exec plans, and a tech-debt tracker.
• Progressive disclosure — agents start with a small stable entry point and are taught where to look next, instead of being overwhelmed up front.
• Application legibility — app is bootable per git worktree; Chrome DevTools Protocol wired into the agent runtime for DOM snapshots, screenshots, and navigation. Logs/metrics/traces queryable via LogQL/PromQL/TraceQL.
• Agent-to-agent review — humans aren't required to review PRs; review is pushed toward agent-to-agent loops (the "Ralph Wiggum Loop").
• Increasing autonomy — L1 (human-in-loop) → L2 (human-on-call) → L3 (autonomous with guardrails), with entropy and garbage-collection strategies for each level.
• Boring tech wins — technologies described as "boring" are easier for agents to model; sometimes cheaper to reimplement a subset than to work around opaque upstream behavior.

Why it over the others

vs Loop Engineering — Loop Engineering is the toolkit (CLIs, scores, patterns) to put this philosophy into practice. This essay is the source — the why and the mental model. Read this first, then reach for the tools.

vs Harness Engineering (field guide) — this is the original OpenAI essay. The field guide is Ryan Lopopolo's companion repo: the same ideas expanded into playbooks, an agent context bundle, and a thesis index you can point an agent at.

vs 12-Factor Agents — 12-Factor is a principles list for production agent systems. This is a case study — what actually happened when one team ran the experiment end-to-end. Complementary: principles + lived experience.

Who it's for

Engineering leaders and staff engineers deciding how to restructure a team around agents. Not a tutorial — a mental model shift.

0 lines of manually-written code. ~1M lines shipped. The leverage point moved from prompts to environments.