Source time Edition date
Close readMicrosoft Research
Formally verifying agent-written full-stack web apps with PDD Kit
用 PDD Kit 形式化验证智能体编写的全栈 Web 应用
PDD Kit shifts review from line-by-line code reading to checking formal specifications, forcing developers to decide whether proved intent outweighs the cost of writing Verus specs for AI-generated full-stack apps.
Judgment
Engineering leads using coding agents must decide whether to require Verus specs for high-risk paths, accepting the spec-writing cost, before letting agents’ code bypass line-by-line review.
Basis
Microsoft Research’s Alex Lavaee and Matthai Philipose say PDD Kit turns intent into a Verus formal spec, proves the agent’s code meets that spec for every input, and falls back to tests where a proof cannot reach. In GitHub Next’s Chopin, they applied it to Pet Store’s login logic: Verus uncovered existing bugs and proved the absence of login bugs with respect to the specification, while the developer reviewed intent and never read generated code line by line. That makes the spec, not line review, the review artifact, but the team must pay to write and maintain Verus specs while agents can ship tens or hundreds of thousands of lines.
Unclear The source does not quantify how much time or effort writing Verus specs adds per feature compared with line-by-line review.
Next Check whether the Pet Store example’s Verus spec and proof are published, so you can estimate spec size and maintenance cost.