Insights

    Agentic AI

    PDD Kit makes the trust boundary the code review

    Devence Lab

    · 2 min read

    Share
    PDD Kit makes the trust boundary the code review
    Photograph · Godfrey Atima / Pexels

    Microsoft Research's login demo shows how teams can prove selected agent-written behaviour while exposing the database, password and token assumptions that still need human review.

    On November 5, Microsoft Research described PDD Kit verifying login logic in a roughly 5,000-line Rust web app. The demonstration is narrow by design. It proves selected behaviour for every input while making the surrounding assumptions visible to a reviewer.

    That distinction matters as coding agents generate larger changes. A plan can say what the developer wants, yet the agent may implement a different rule and then write tests that pass against its own implementation. PDD Kit turns reviewed intent into Verus specifications, proofs for the parts the verifier can inspect, and tests for the parts it cannot.

    Specify both sides of a promise

    The login request initially said that a token should be issued only with the right password. Microsoft points out that a service which rejects every login satisfies that statement. The team had to add the converse: a correct password should work, subject to an operational failure. This small clarification changes what the proof must establish.

    For an engineering team, the lesson is to review expected successes and refusals before an agent writes code. Authentication, approval and payment rules often hide the same one-sided promise. A verifier can check the agreed statement exhaustively; it cannot decide whether the statement captures the business rule.

    The most valuable review is the decision about what the proof may assume.

    Name every trusted component

    In Microsoft's example, Verus proves the login flow against its specification. It does not prove that SQLite returns the correct row, bcrypt checks the right hash, the clock reports the right time or token signing behaves as assumed. PDD Kit records these as trusted contracts and attaches tests to their clauses.

    The distinction is operational. A green proof says the code meets the stated contract if those trusted components behave as described. It says nothing about an omitted requirement, a hung database call or an unseen failure in a dependency. Microsoft explicitly lists those limits in the plan instead of hiding them behind a single verification badge.

    Move the release gate to intent and evidence

    Use the approach first on a small, high-consequence path. Ask reviewers to approve the plain-language intent, including edge cases and failure behaviour. Identify which functions are proved, which are tested under assumptions, and which remain outside the claim. Require each trusted contract to have an owner and an executable check.

    PDD Kit is a research demonstration, not evidence that an entire generated application is correct. Its useful contribution is a review record that separates guarantees from trust. Teams adopting agent-written code can copy that discipline even when their stack cannot use Verus: an explicit contract and a visible boundary make a release decision more defensible than a large set of agent-written passing tests.

    Sources

    1. Formally verifying agent-written full-stack web apps with PDD Kit · Microsoft Research

    Written by the Devence Lab research team.

    Share