Microsoft Research introduced on November 5 the PDD Kit (Proof-Driven Development Kit), a toolkit that uses mathematical proofs to check whether code written by artificial intelligence agents actually does what the developer asked for. The proposal, described by researchers Alex Lavaee and Matthai Philipose on the official Microsoft Research blog, shifts the focus of review: instead of reading the AI-generated code line by line, the team reviews the intent — and a machine proves the code fulfils it.
The problem PDD Kit tries to solve is one of scale. Coding agents already ship whole features in one sitting, spanning tens or hundreds of thousands of lines, and reading every line of a pull request no longer scales. The natural-language plan is supposed to work as a contract between the developer and the agent, but nothing forces the agent to follow it — and the tests it writes itself tend to validate what it built, not what was requested.
How it works
PDD Kit turns intent expressed in plain language into a formal specification in Verus, a proof language for Rust co-developed by Microsoft Research, Carnegie Mellon University, ETH Zurich and other institutions. The verifier then proves the agent's code meets that specification for every possible input — and falls back to tests where a proof cannot reach. The kit runs inside Chopin, GitHub Next's real-time, multiplayer planning environment.
The workflow has two halves: first, the team reviews intent in plain language, with PDD Kit asking the questions a careful reviewer would ask and recording the answers in an intent artifact. Then the kit generates proofs that the code satisfies the formal specifications, plus quality tests for the rest. No formal-methods background is needed to use the process.
The test: the login of a Rust storefront
To demonstrate the kit, the team built Pet Store, a full-stack web app written entirely in Rust, with about 5,000 lines of code: the back end uses Axum and SQLite and handles inventory, promotions, orders and checkout, while the storefront is a Leptos app compiled to WebAssembly. The exercise focused on the login logic, starting from a request of just two sentences: a token is issued only with the right password, and only administrator accounts get administrator tokens.
Verus formally verified five functions — login_flow, login_core, lookup_target, decide_login and issue_token — and, according to the researchers, hardened the code and uncovered existing bugs, proving the absence of login bugs with respect to the specification. The developer reviewed only the intent and never read the generated code line by line.
What the proof covers — and what it leaves out
Microsoft is careful to delimit the guarantee. Six components sit outside the proof, behind assumed contracts the developer accepts one by one: the two database lookups, password checking with bcrypt, the clock, token signing, database error handling and unpacking the request. Each of them is covered by tests.
The proof also says nothing about what is not in the specification: whether login will answer, how fast, what the attempt limit is, or whether the error message is clear. And PDD Kit only verifies Rust today — the language Verus works on. Projects in other languages can be ported to Rust or call verified Rust through bindings.
Against “vibe verification”
One of the kit's features is the mutation check: pdd-prove plants one bug per specification clause in a scratch copy of the code, and Verus has to reject each mutant. If a mutant gets through, the clause constrains nothing — and comes back as a gap. It is the defence against what the researchers call “vibe verification”: weak invariants that verify but guarantee almost nothing.
When the tasks pass, the kit opens pull requests with the verified code and marks each requirement as formally verified or empirically checked. The pdd-ship workflow adds the Verus check to continuous integration, so any later change to login has to keep proving against the same specification.
PDD Kit is a research prototype, with access described as a limited research preview. Even so, it points in a concrete direction for a real problem: as agents write more and more code, reviewing intent — and proving it was honoured — may be the only way review keeps up with the pace.
Sources
- Microsoft Research — Formally verifying agent-written full-stack web apps with PDD Kit (November 5, 2026)
- The Agent Times — Microsoft Research Describes PDD Kit for Agent-Written Code
The official @MSFTResearch account also shared the work on X on November 5, 2026.
Image credit: Microsoft Research — PDD Kit decision flowchart, the featured image of the official article on the Microsoft Research blog.
Transparency: this article was produced with the help of artificial intelligence, cross-checked against public posts on X and sources on the internet. Check the original sources to confirm the information.
By Muse
