Vaibhav Gupta is the co-founder and chief executive of Boundary, where he develops BAML, an open-source programming language for dependable AI applications and agents. He believes software-writing agents require stronger architectural guarantees, clearer execution traces, and programming tools that remain trustworthy when engineers cannot inspect every generated line.
From mixed reality to BAML
Gupta worked on Microsoft HoloLens and mixed-reality systems before joining Google, where his projects included augmented reality, on-device performance, and the Pixel 4’s depth and face-unlock systems. He subsequently worked on infrastructure at D. E. Shaw. His longstanding interest in programming tools is visible in an early online compiler.
With Aaron Villalpando, Gupta first built Gloo, a voice-first collaboration product, and joined Y Combinator’s Winter 2023 batch. The founders explored API infrastructure, automated drive-through ordering, embeddings, classification, and information extraction before beginning BAML in August 2023. Gupta’s account of Boundary’s pivots traces the shift toward a typed language that integrates language models into production software without brittle prompt strings, opaque abstractions, or vendor-dependent interfaces.
A programming language for AI agents: Gupta treats prompts, model selection, typed outputs, testing, and streaming as language-design problems. His argument for agent-oriented syntax envisions AI-specific constructs that remain interoperable with established programming languages.
Schema-aligned parsing: BAML extracts typed data from imperfect model outputs without requiring rigid JSON generation or a single provider’s function-calling interface. Gupta’s structured-output analysis examines tradeoffs involving portability, latency, token consumption, and accuracy.
Enforceable architectural invariants:Boundary replaces conventional code reviews with compact architecture documents, versioned design proposals, Slack-integrated collaboration, dependency visualization, and automated checks. Engineers can choose different coding assistants while shared constraints prevent generated code from violating architectural boundaries. Gupta’s GitHub-repository mutex joke captures the coordination problems created by concurrent human and agent contributions.
Execution-based agent evaluation: Agents generate BAML programs, inspect execution histories, identify potential defects, and propose repairs; humans distinguish genuine problems from hallucinations. Comparing tool calls, errors, and outcomes lets Boundary test language features empirically instead of relying on intuition.
Compiler-checked trust: Gupta advocates execution tracing, agent-friendly code discovery, inferred error types, exhaustive error handling, and typed interoperability with languages including Python. His AI Engineer World’s Fair presentation demonstrates how architectural visibility and compiler guarantees can make agent-generated systems understandable without demanding wholesale application rewrites.
How do you build a reliable programming language without code reviews? Vaibhav Gupta’s answer combines architectural constraints, agent-driven testing, and tooling that makes behavior easier to inspect.
