Breaking

Boris Cherny Says Claude Opus 5.5 Verification Run Produced 16 Claude Agent SDK PRs

September 23, 2026 at 20:26 EDT

Software Dev & Coding
AI Agents
Foundation Models

Boris Cherny, creator of Claude Code at Anthropic, said Claude Opus 5.5 was used to formally model parts of the Claude Agent SDK with Lean, turning a short prompting session into 16 pull requests that fixed bugs and race conditions. The report is a timely test case for whether frontier coding agents can make formal methods useful in ordinary software maintenance, not only in research settings.

Sep. 22, 2026 · Anthropic

Formal verification moves into the agent workflow

Claude Opus 5.5 was used with Lean and TLA+ to examine the Claude Agent SDK, producing 16 pull requests aimed at bugs and race conditions.

pull requests from the run

input price per million tokens

$20

output price per million tokens

token context window

The reported output, block by block

16 PRs

Each segment represents one pull request produced from the formal-verification pass, aimed at bugs and race conditions in SDK behavior.

How the loop works

Model SDK logicRepresent state and behavior formally

→

Check proofsLet Lean or TLA+ reject invalid claims

→

Patch failuresTurn verified pressure into reviewable code

Why developers care

The workflow could lower the barrier to formal methods and catch bugs that ordinary tests or code review may miss.

What still needs scrutiny

A proof checks the stated model, so maintainers must still confirm that the model matches the real SDK code.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.