{"slug": "agentic-proof-oriented-programming", "title": "Agentic Proof-Oriented Programming", "summary": "As of late 2025, AI agents integrated into coding environments such as Claude Code, OpenAI Codex, and GitHub Copilot CLI have made automated formal proofs a reality, according to researchers who have used agents to author over 100,000 lines of verified F* and Pulse code, including a generational garbage collector for OCaml 4.14.0 and a TLS-1.3 implementation interoperating with OpenSSL. The researchers, led by the F* project, emphasize that while agents can produce software at a scale exceeding human review capacity, proof checkers like F* ensure correctness against formal specifications, but humans must carefully design abstractions to validate intent.", "body_md": "# Agentic Proof-Oriented Programming\n\nAI-based automation of formal proofs has received a lot of attention in the past few years, with lots of promise but few real successes. However, starting in late 2025, with the availability of models such as Claude Opus 4.5 and GPT-5.2 (and beyond), and especially their integration within agentic coding environments such as Claude Code, OpenAI Codex, and GitHub Copilot CLI, the promise of AI automation for proofs has largely become a reality. It is now possible to have agents author hundreds of thousands of lines of programs and proofs, including in F* and Pulse, but also in other proof assistants including Lean, Coq, Verus, Dafny, and others.\n\nFor example, in a series of [blog posts](https://risemsr.github.io/blog) we have reported on\nour ability to use AI agents to automatically formalize the classic “CLRS”\n[Introduction to Algorithms](https://en.wikipedia.org/wiki/Introduction_to_Algorithms)\ntextbook in F* and Pulse, in just a few weeks of agentic coding.\nBeyond just textbook algorithms, we have also been able to use agents to build a\nformally [verified generational garbage collector for OCaml](https://github.com/FStarLang/pulse-verified-gc)\nwith proofs of correctness comprising more than 100,000 lines of F* and Pulse, authored\nentirely by agents. The resulting verified, executable C code can be used as a\ndrop-in replacement for the OCaml 4.14.0 garbage collector, and has been tested to work with\nthe OCaml 4.14.0 compiler and runtime system.\nGoing further, we have also been able to use agents to build a verified implementation of\nthe [TLS-1.3 protocol](https://github.com/project-everest/mitls-fstar/tree/agentic),\nwith implementation and proofs again exceeding 100,000 lines of F* and Pulse.\nThe resulting C code interoperates with OpenSSL to negotiate TLS connections and\nsend application traffic.\nOthers have reported similar successes, using agents to formalize proofs of [programming\nlanguage metatheory](https://proofsandintuitions.net/2026/03/18/move-borrow-checker-lean/),\nparts of [verified compilers](https://zoep.github.io/blog/2026/04/17/machine-generated-code-machine-checked-proofs/),\nor porting [C code into Lean with proofs of correctness](https://github.com/kim-em/lean-zip).\n\nThat said, agentic proof-oriented programming is an extremely new paradigm, the agents are evolving rapidly and there are many open questions about how best to structure interactions between humans and agents to achieve the best results. What is clear is that agents can produce software artifacts at a scale which, if not structured correctly, can far exceed the ability of humans to review and validate the results. The promise of proof-oriented programming in this context is that proof checkers like F* can be used to ensures that the software matches a formal specification, but one still has to be able to carefully check that what is proven matches what is intended.\n\nIn this part of the book, we offer a few principles and techniques for agentic proof-oriented programming. Thankfully, a lot of what we’ve come to know as good proof-engineering practice is still very much relevant in this new setting. Our main message is that humans need to think carefully about the right abstractions for designing a system so that agents are effective at producing proof-oriented software, and, perhaps more importanly, that humans are able to understand the results.", "url": "https://wpnews.pro/news/agentic-proof-oriented-programming", "canonical_source": "https://fstar-lang.org/tutorial/book/agentic/agentic.html#agentic-proof-oriented-programming", "published_at": "2026-08-13 02:13:58+00:00", "updated_at": "2026-08-13 02:40:49.999162+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-research", "ai-tools", "ai-products"], "entities": ["Claude Opus 4.5", "GPT-5.2", "Claude Code", "OpenAI Codex", "GitHub Copilot CLI", "F*", "Pulse", "OCaml"], "alternates": {"html": "https://wpnews.pro/news/agentic-proof-oriented-programming", "markdown": "https://wpnews.pro/news/agentic-proof-oriented-programming.md", "text": "https://wpnews.pro/news/agentic-proof-oriented-programming.txt", "jsonld": "https://wpnews.pro/news/agentic-proof-oriented-programming.jsonld"}}