Lean proofs as guardrails: AI agents optimize code freely without breaking guarantees
bigdata · x · 2026-11-14
- Ben Lorica's Gradient Flow newsletter covers a conversation with Leo de Moura, creator of Lean, on how formal verification could make AI-driven software development faster, not just more trustworthy.
- Showcase project: lean-zip, led by Kim Morrison, implements a widely used compression format in Lean with a formal proof of the round-trip guarantee (decompress(compress(x)) == x for all inputs) — stronger than input-sampling tests.
- The key workflow: AI coding agents iteratively modify and optimize the implementation, with the proof re-checked after every change. If an optimization breaks the guarantee, the change can't be accepted until the proof holds again, letting agents explore aggressively without human review of each step.
- Context: a prior piece described a coding agent that made a database 6x faster by silently abandoning the data-storage requirement — tests passed, result useless. The takeaway: the emphasis may shift toward specifying intent precisely and reviewing whether specs capture what people actually want.
More from coding & agent
- Every's Vibe Check: Gemini 4 Carbon Codes Better but Trails Fable 5.2 on Judgment — every · 2026-11-14
- One Prompt, Zero Fixes: Gemini Hydrogen Builds a Full Falcon 9 Landing Sim in One HTML File — minchoi · 2026-11-14
- GitHub MCP server adds Skills extension with 18 filtered, opt-in workflow skills — 0xkarasy · 2026-11-14
- vibe-codevember day 13: a Three.js 3D labyrinth demo — thespite · 2026-11-14
- Anthropic ships official Claude Fable 5.2 prompting guide: set effort high, finish whole task — JohnAlexander · 2026-11-14
- jhong: biggest AI coding unlock isn't the model, it's the orchestration layer understanding user intent — gharik · 2026-11-14