AI Agents Need a New Syntax (boundaryml.com)

Vaibhav Gupta announced BAML (Basic-Ass Machine Learning), a new open-source (Apache 2) syntax and toolchain for building AI applications and agents that treats prompts as first-class functions. Motivated by the same maintainability problems React and Tailwind solved for web dev, BAML aims to move developers away from brittle string-heavy prompt engineering toward a composable, language-native workflow: every prompt is a function with typed parameters, a declared model, and an explicit prompt body. That lets standard LSP/IDE tooling locate, version, and hot-reload prompts, supports one-line model swaps, static retries/fallbacks, and dramatically speeds iteration (compiler ~30 ms). Technically, BAML compiles to native code in multiple languages (Python, TypeScript, Ruby, etc.) via a shared Rust core, or exposes a local REST/OpenAPI for unsupported languages (C++, Java, Go, …). It includes SAP (schema-aligned parsing) to produce flexible, structured outputs (e.g., markdown inside JSON) without relying on tool-use or function-calling APIs, and a streaming-aware type system (supporting Partial<T>[] semantics for incremental results). Privacy/offline-first design restricts network calls to explicit model endpoints only. For AI/ML engineers, BAML promises maintainability, cross-language integration, type-safe streaming, and faster prompt iteration—practical infrastructure for the next wave of agentic systems.