Nobody reads code anymore: 6 field-tested rules for shipping safely with AI
Specialist_Agent3599 · reddit · 2026-09-03
After two months of record shipping speed, the author's team stopped reading AI-generated code entirely. Instead of fighting it, he shares 6 field-tested verification practices: write far more tests than feel necessary; never review code with the model that wrote it (NVIDIA's fine-tuned 8B reviewer beat 70B/340B models; a 7B HF reviewer beat a 70B baseline by 46%); prefer several cheap independent review passes over one perfect pass (a 12B reviewer experiment confirms it); never trust green tests — LLM suites have hit 100% coverage with only a 4% mutation score, so break code on purpose; run a dedicated security pass (Veracode: 45% of generated code fails security tests, 86% of XSS cases insecure); and tier human review by blast radius — auth, payments, and migrations always get human eyes, since data from 22,000 developers shows review time up 441% and unreviewed merges up 31% under heavy AI adoption.
More from coding & agent
- JPMorgan, BofA and Wells Fargo Block Uncontrolled AI Agents From Customer Bank Accounts — emmanuelvivier · 2026-10-25
- Code is clay now: an iOS developer on how agentic coding reshapes software building — bibryam · 2026-10-25
- Row-Bot 5.3 ships embedded browser, tiered approvals and 40% lighter context for local-first agent — Acceptable-Object390 · 2026-10-25
- Task-specific inference engines emerging as specialization costs collapse, argue researchers — _AustinCalvert_ · 2026-10-25
- Daniel Miessler on models vs harnesses: better models mean less harness — _AustinCalvert_ · 2026-10-25
- Celesto: open-source secure persistent computer for AI agents — aniketmaurya · 2026-10-25