Formally Verifying the Claude Agent SDK with Opus 5.5 Yields 16 Bug-Fix PRs
LingmingZhang · x · 2026-09-24
Boris Cherny used Claude Opus 5.5 with Lean to formally verify the Claude Agent SDK, generating 16 PRs fixing bugs and race conditions from a couple of short prompts; he also combines Lean with TLA+ to catch data-flow, concurrency, and state-management issues, arguing the approach finds bugs humans would likely miss.
Formal verification researcher Grigore Rosu went further, sketching "Intent Computing": AI generates code, spec, and proof together from human intent; the spec is rendered back in natural language for approval; and an independently checkable proof guarantees the artifact matches the intent. In his framing, AI is the new computer and intent is the new programming language.
Related event: Opus 5.5 with Lean formally verifies Agent SDK(18 posts)→
More from coding & agent
- CC Switch dev found stars flatlined after moving downloads off GitHub — tinyfool · 2026-09-24
- Sentry CEO: agents built by agents are the worst code models generate — zeeg · 2026-09-24
- GitHub partners with Muse to review PRs, issues and notifications without switching tabs — unixterminal · 2026-09-24
- Dev take: the most fun projects are ones where code details don't matter — kieranklaassen · 2026-09-24
- simonw talks to his blog's Datasette via MCP using ChatGPT Voice on iPhone — reach_vb · 2026-09-24
- Designer builds a tarot app in 2 hours with Figma, Claude and Midjourney — AlexisDC · 2026-09-24