Opus 5.5 reportedly produces 16 bug-fixing pull requests for Claude Agent SDK
A user credits formal verification with Lean for the fixes. An IDS paper coauthor reports that agents developing code and proofs together achieved 3x Claude Code’s success rate on distributed systems specifications.
TLDR
A user says a couple of short prompts to Opus 5.5 yielded 16 pull requests fixing bugs and race conditions while formally verifying the Claude Agent SDK with Lean. They also describe sometimes combining Lean and TLA+ to find data-flow, concurrency and state-management issues, while acknowledging limited familiarity with either language.
Quoting that account, an IDS paper coauthor says their paper was accepted for a NeurIPS oral presentation. They report that agents co-evolving code and proofs achieved 3x Claude Code’s success rate on distributed systems specifications, and point to further work on ensuring specifications capture human intent.
