AutoCLRS formalizes CLRS with agents
Microsoft Research’s RiSE team says AI agents, guided mostly by natural-language feedback, produced more than 100,000 lines of F* and Pulse specifications, code, and proofs covering roughly 50 classic CLRS algorithms and data structures. The notable part is not just scale, but that the project is public as AutoCLRS, making agent-driven formal verification inspectable instead of purely aspirational.
This is one of the clearest signs that agentic coding is starting to matter in formal methods, not just app scaffolding. AutoCLRS is less a polished library release than a live demonstration that proof engineering can be accelerated dramatically while shifting the real bottleneck to spec review.
- –The project covers a surprisingly broad chunk of CLRS, including sorting, graph algorithms, dynamic programming, and data structures, which makes it more convincing than a cherry-picked demo
- –The post is honest about where agents still fail: weak or distorted specifications can still produce mechanically checked but practically wrong results
- –The repo turns formal verification into something other teams can inspect, reuse, and benchmark rather than treating “AI-generated proofs” as a black-box claim
- –The addition of termination proofs in Pulse during the effort shows a useful feedback loop where agents do not just use tooling, they expose missing capabilities in the tooling itself
DISCOVERED
31d ago
2026-03-11
PUBLISHED
32d ago
2026-03-10
RELEVANCE