Vibe coding produces 50K-line PRs, but nobody can verify they actually work
srchvrs · x · 2026-09-18
In a thread, srchvrs pushes back on the current AI coding narrative: with Codex and Claude Code, developers are landing 10K–50K-line PRs and code output has exploded, with Boris Cherny, Dario Amodei, and Sam Altman all cheering.
The catch, he argues: nobody and nothing can actually verify these giant PRs work.
- Testing pipelines are dominated by vibe-coded tests, often written by the same model that generated the implementation — grading its own homework
- Serious integration testing, production-like testing, end-to-end verification? "Maybe eventually."
- The deployed "miracle package" only works as long as someone constantly hand-holds it: live debugging, on-the-fly patches, and handling surprises once the code meets reality
More from coding & agent
- AgentSky launches as an agent marketplace: 40+ coding agents in browser, 44x cost gap between models — Scobleizer · 2026-09-18
- MiniMax open-sources its Code CLI v0.4.12, claims SOTA on FrontierHarness Eval — MiniMax_AI · 2026-09-18
- MiniMax Code CLI is now open, inviting exploration of new agent paradigms — MiniMax_AI · 2026-09-18
- AWS Ships Six Open-Source Skills to Let Coding Agents Deploy Hugging Face Models on SageMaker — AWS ML Blog · 2026-09-18
- Three guys with Claude and Codex subscriptions chained 6 vulns into an advanced attack — xennygrimmato_ · 2026-09-18
- Outerloop 0.2.0 ships persistent agent states and a unified message inbox — mengyer · 2026-09-18