PDD Kit proves AI-agent code with Verus on Pet Store login

The premise behind PDD Kit is that the plan has become the contract between a developer and a coding agent, yet nothing forces the agent to follow it. An agent can implement something slightly different and then write tests that pass for its own implementation, so continuous integration fills with tests while the gap between what was asked for and what shipped stays hidden until a bug report surfaces. Microsoft Research argues tests alone can't close that gap because they exercise only a tiny slice of what a program can do, which is where proofs come in.

The demonstration shows how the toolkit handles the ambiguity an agent would otherwise settle silently. In Pet Store's login, the sentence "you only get a token with the right password" rules out a wrong password getting a token, but a login that refuses everyone keeps that promise perfectly. PDD Kit surfaced that the converse, "the right password always works," was the half that rules it out, and the developer chose both directions, making the requirement an if-and-only-if. The two readings differ in a single cell, a right password that gets no token.

Where proof stops, PDD Kit makes the line explicit. For login, the developer signed off on six trusted pieces, including the two database lookups, check_password around bcrypt, generate_token, expiry_from_now and the handler shell, with each assumed clause getting a planned test written before the test exists. The toolkit also records what nobody is promising, such as whether the login ever completes or how other endpoints read the role.