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)→

Original post →

More from coding & agent

coding & agent channel →