Opus 5.5 Formally Verified Claude Agent SDK in Lean, Producing 16 Bug-Fix PRs
_catwu · x · 2026-09-23
Boris Cherny used Opus 5.5 to formally verify the Claude Agent SDK with Lean — a couple of short prompts yielded 16 PRs fixing bugs and race conditions. He also combines Lean with TLA+ to probe data flow, concurrency, and state management, noting Claude excels at both despite his own limited fluency. He argues formal modeling surfaces bugs humans would likely miss and asks if formal verification is the future of bug finding.
More from coding & agent
- Syrupy: the popular pytest snapshot plugin keeps tests of big outputs readable — KhuyenTran16 · 2026-09-23
- Tame bloated pytest assertions with Syrupy snapshot files — KhuyenTran16 · 2026-09-23
- Claude Opus 5.5 lands on Netlify AI Gateway and Agent Runners with zero configuration — thisiskp_ · 2026-09-23
- Stop chasing phantom LLM eval swings: bootstrap confidence intervals in 20 lines of Python — bgoncalves · 2026-09-23
- Jev API explodes at $0.042/M tokens: a hands-on checklist from desktop agents to drone control — blaizedsouza · 2026-09-23
- The prompt to run before wiring an agent to another service — gethackteam · 2026-09-23