Summary
At the AI Engineer conference, Vaibhav Gupta — co-founder of Boundary and creator of the BAML programming language — presents a counterintuitive set of engineering practices his team developed over three years of shipping a production compiler with no code reviews and no standardization on AI tooling. The central concept: “slop” (any code the team doesn’t actually read) is inevitable, so the answer is to build invariants into tooling and architecture that catch AI-generated mistakes automatically rather than trying to prevent agents from writing messy code in the first place.
Gupta walks through four concrete artifacts his team built: a minimal `architecture.md` file encoding only stable, multi-year invariants that any model can parse; a custom design doc tool with Slack integration that turned design reviews into a spontaneous social channel; a dependency graph visualizer with CLI-enforced semantic boundaries that catches architectural violations in CI before they merge; and a mandatory peer-review gate on design docs (not code) that organically drove quality up without policing how engineers used AI.
The talk’s most technically ambitious segment introduces a visual code navigation interface built on top of BAML that lets engineers — and agents — explore execution traces and semantic boundaries without reading raw source. Gupta argues that as codebases become predominantly AI-generated, the fundamental interface for understanding software must change, and that first-principles language design is more durable than layering quality controls on top of JavaScript or Python.
📺 Source: AI Engineer · Published July 31, 2026
🏷️ Format: Hands On Build
