Skip navigation
How do you solve the trust problem with AI-generated code?
0:00 / 1:24

How do you solve the trust problem with AI-generated code?

Microsoft Research

@MSFTResearch557K followers

Visit on X
192 likesDownloadOriginal post

21,201 views

How do you solve the trust problem with AI-generated code? AI agents can generate code at a scale too great for humans to review line-by-line. That makes it difficult to ensure the code will do what the developer intended, and it leaves the final output vulnerable to bugs and security risks. Proof-Driven Development Kit turns plain-language developer intent into a specification in Verus, applying formal methods traditionally used for safety-critical systems. PDD Kit uses mathematical proofs to test agent-generated code against the spec where possible, and clearly indicates where the proofs can’t reach. The result is more reliable agentic code. Researchers Matthai Philipose and Alex Lavaee break down Proof-Driven Development in the MSR blog:

14 Comments

Comments from the original post on X (Twitter)

  • Raúl Nogales 🇪🇸🇺🇸's profile picture

    @raulnogales

    The trust problem is not that agents write bad code. It is that they write plausible code at a scale no human can review line by line. The fix is not more review. It is tests, constraints and code the agent can prove, not just produce.

    3 likes

  • Mykyta Pavlenko's profile picture

    @mktpavlenko

    Proof moves the review from the code to the Verus spec. Fewer lines to read, but each one now carries the whole intent.

    2 likes

  • RAZA | AI EXPLORER's profile picture

    @razaaitech

    Formal verification could be the missing trust layer for AI-generated code. Proving correctness beats simply hoping for it!

    1 like

  • Jacob Liu's profile picture

    @opendocyai

    At that scale, review has to move from reading every line to checking evidence: tests passed, rules satisfied, and a named person accountable for what ships. Reviewers verify the record, not the whole output.

    1 like

  • Krrish's profile picture

    @KrrishKapoor8

    do the unproven parts still merge, or get blocked?

    1 like

  • IdentArk's profile picture

    @identArk

    A coding agent may give correct answer based on failed reasoning, how does PDD kit handle this reasoning failure

  • Jon Cheney's profile picture

    @cheneypiano

    When agents generate more code than anyone can review line-by-line, what's the review gate you're trusting instead?

  • toby's profile picture

    @Toby___l

    @bot 介绍一下这个

  • Amelia | AI & Tech's profile picture

    @AmeliaonAI

    Proofs checking agent code could finally build real trust

  • Code Particle's profile picture

    @CodeParticle

    The part worth copying even without formal methods: mark where verification can't reach. Review of agent code tends to fail quietly at exactly that boundary. A spec that labels its own unverified edges tells the reviewer where to spend the hour.

  • Prasen Shelar's profile picture

    @PRaSeN_S

    That list of unproven parts could go straight onto the pull request, so a security reviewer spends their time there and a named person signs off before merge. Does the kit output it in a form a CI check can read?

  • Daniel Saadia's profile picture

    @ddyods

    Yes ! Reading every line does not scale. What I do is lock an expect.yml before the agent starts, then a script checks the real diff against it.

  • RemoteBrowser's profile picture

    @RemoteBrewser

    The scale problem is real, but line-by-line review was never the only gate. For agent-generated code, the leverage is in executable checks: tests the agent must pass, plus actually running the output in a sandbox and watching behavior. Reviewing intent beats reviewing diffs.

  • Veyra | AI & Tech's profile picture

    @VeyraWithAI

    Proofs over vibes is exactly what agentic code needs