Loop engineering: gate the spec, withhold the grader, label the jury
Skip the prompt engineering. The agent runs, fails, and gets smarter every generation. Loop engineering is our name for the discipline of building that loop so the growth has a boundary: the grading command and expected result never make it into the success contract we hand it.
This page documents how Ouroboros does it, with the code to check. It is a practice description; we are not claiming a market-wide definition.
Three principles, each greppable
- The loop starts gated. An interview scores how vague the request is, and a score above 0.2 blocks spec generation until you pass force explicitly. The gate makes vagueness a decision instead of an accident.
- The grader stays outside the loop. When an acceptance criterion defines a verify command or an expected-output assertion, those values are omitted from the worker's contract block, with no flag to put them back. Redaction elsewhere is literal over five encodings, so a reshaped copy can still get through; that gap is tracked in a public issue.
- The jury is labeled, not assumed. Evaluation runs carry an independence state; same-vendor voters are filtered, and if that would leave fewer than two, the original roster is rolled back rather than pretending a one-voter jury is a jury.
Every claim above has a file:line address: see the source tour page for the exact coordinates.
The anthology
Pieces we have published on the practice, in the language they were written in:
- English: Your agent loop is teaching the model to cheat (dev.to), An ambiguity gate for AI coding agents (dev.to)
- 한국어: 당신의 에이전트 루프는 모델에게 커닝을 가르치고 있다 (dev.to)
- 日本語: Qiita edition
- 中文: X thread, 4 posts. 如果说 harness 是执行骨架，那么 loop engineering 说的是让这副骨架一代比一代聪明的边界规范。
- Hands-on: the learn guides (en / ko / zh) walk from install to a first interview.
Why a page instead of more posts
The term is empty where we checked: marketplace exact-match came back zero when we measured (2026-08-11). A single page that defines, shows the code, and collects the pieces is what search engines, citers, and agents can actually land on. Posts age; this page is maintained.
If your loop uses a different spec gate, grader boundary, or jury rule, send one public implementation and name the boundary it addresses. We will add verified examples to this anthology, including designs that disagree with ours.