Reviewing every line of code an AI agent writes no longer scales. Microsoft Research has presented PDD Kit, a tool that turns a developer's intent into a formal specification and uses a mathematical proof to check that the code an agent generates satisfies it for every possible input, without reading it line by line.
The problem is concrete. Coding agents now deliver entire features in a single session, often tens or hundreds of thousands of lines. A natural-language plan is supposed to be the contract between developer and agent, but nothing forces the agent to follow it. The agent can implement something slightly different and write tests that pass on its own code rather than the code requested. Test coverage looks healthy, and the gap between what was asked and what shipped stays hidden until a user reports a bug.

Image: itai-gendler — https://github.com/itai-gendler/pdd-kit
Review asks whether something looks right; a proof asks whether it is right. A proof covers every input, including cases a test suite never reaches. Until recently, proofs were too costly for everyday software, requiring specialists, unfamiliar languages and proofs far longer than the code. Two developments have changed that: agents can now help generate proofs from well-defined specifications, and they can help write the specifications too. Microsoft calls this approach Proof-Driven Development, and PDD Kit implements it.
How it works
PDD Kit runs inside GitHub Next's Chopin, a real-time, multiplayer planning environment where teams and agents work in one shared document. The workflow has two halves. First, the team settles intent in plain language: the tool asks the questions a careful reviewer would ask, explains the existing code and records the answers in an intent document that links intent to the formal specification or tests. The part of the intent that can be formalised becomes a formal specification; the rest, which must be taken on trust, becomes an informal specification for tests. Second, PDD Kit generates proofs that the code satisfies the formal specifications and high-quality tests for the informal ones.
Every result states which kind of verification supports it. Formal verification is a mathematical proof checked by a verifier, which in PDD Kit's case is Verus, a proof language for Rust co-developed by Microsoft Research and several universities. Empirical verification runs the code and observes its behaviour, covering what a proof cannot reach, such as a third-party cryptography library or a user interface. A passing test only shows the code worked on the inputs tried, so PDD Kit labels it as evidence rather than proof.
The dividing line is the Trusted Computing Base: the code the proof assumes behaves as described. PDD Kit makes that line explicit and asks the developer to sign off on each piece.
The Pet Store case
To show integration with existing codebases, the team built Pet Store, a full-stack web app written entirely in Rust, with an Axum and SQLite backend that handles inventory, promotions, orders and checkout, and a Leptos storefront. In total, about 5,000 lines of code.

Image: itai-gendler — https://github.com/itai-gendler/pdd-kit
The test focused on login logic, because a bug there can hand an attacker a customer's account or an admin's controls. The developer's request ran to two sentences: a token only with the correct password, and admin tokens only for admin accounts. The team stresses it kept the request deliberately rough.
The result: Verus hardened the code and uncovered bugs that already existed. The developer reviewed the intent without reading the generated code line by line. The post is clear about the limits: the proof covers login against the written specification, not the entire application. The team recommends starting with the most sensitive parts and extending verification incrementally.
Availability
According to a Japanese report, PDD Kit is limited to Rust code and offered as a limited research preview. Microsoft's blog does not state wider release timing or terms of use, and does not say how long generating and maintaining proofs takes on larger applications.
The announcement's core idea is to shift trust from manual line-by-line review to a specification that a human approves. Whether the approach holds up on systems far larger than Pet Store remains to be seen.

No comments yet. Be the first!