Proof-Oriented Engineering: defining 'good' before you let an agent crank
gregmushen · x · 2026-09-25
The author proposes "Proof-Oriented Engineering": the promise of modern agent harnesses is set a goal and come back hours later with everything done, but the hard part is articulating what 'good' looks like. He borrows Justice Stewart's 1964 line on obscenity — "I know it when I see it" — to show most acceptance criteria resist clear definition.
This is both the joy and frustration of LLMs: articulate requirements well and it's magic; be off and it drifts. The article argues for organizing agent engineering workflows around provable acceptance criteria.
More from coding & agent
- Two Prompts, One Showreel: Claude Code Makes a Full Motion Graphics Video — tristanbob · 2026-09-26
- Non-black-box embeddings: use an LLM's answers to named questions as your search vectors — kieranklaassen · 2026-09-26
- OpenAI case study: Proaction boosts sales 60%, saves 75+ hours with Codex — OpenAI News · 2026-09-26
- Vercel Wins 86 of 86 Frontend Decisions as Claude Code's Default Deploy Target — Agents Have No Procurement Cycle — le_james94 · 2026-09-26
- Testing Jev as a confidence-scored router to offload small LLM calls in oncology agents — Ubunta · 2026-09-26
- Kitaru adds transition failure matrices to its MCP server to debug agent failures — strickvl · 2026-09-26