ZeroLang — The Programming Language Designed for AI Agents

Introduction

ZeroLang is a programming language created by Vercel Labs specifically for building and orchestrating AI agents. Rather than retrofitting agent patterns onto existing languages, ZeroLang treats tool invocation, sandboxed execution, and structured data flow as first-class language constructs. Written in C for maximum portability, it compiles to a lightweight runtime that can be embedded anywhere.

What ZeroLang Does

• Provides language-level primitives for agent tool calls and structured output
• Runs agent programs in a sandboxed environment with controlled resource access
• Supports deterministic workflows with built-in retry and error-handling semantics
• Compiles to a minimal runtime embeddable in servers, CLIs, and edge functions
• Offers a concise syntax optimized for describing multi-step agent logic

Architecture Overview

ZeroLang compiles source files into bytecode executed by a lightweight virtual machine written in C. The VM provides built-in opcodes for tool dispatch, parallel execution, and structured I/O. A capability-based sandbox restricts what each agent program can access at runtime, making it safe to run untrusted agent code. The compiler is a single static binary with zero external dependencies.

Self-Hosting & Configuration

• Install the single-binary compiler on Linux, macOS, or Windows
• Write agent programs in .zero files using the ZeroLang syntax
• Configure tool registries to connect agents to external services
• Embed the ZeroLang runtime in existing applications via the C API
• Set sandbox policies to control file, network, and process access

Key Features

• First-class tool-use syntax eliminates boilerplate in agent code
• Capability-based sandboxing for safe execution of agent workflows
• Zero-dependency single binary compiles and runs on any platform
• Deterministic execution model simplifies debugging and testing
• Designed for embedding in serverless and edge environments

Comparison with Similar Tools

• Python — general-purpose but lacks built-in agent primitives and sandboxing
• TypeScript — widely used for agents but requires heavy framework layers
• Starlark — sandboxed but designed for build configs, not agent workflows
• Jsonnet — configuration language without imperative agent control flow
• Dagger (CUE) — pipeline-oriented, not designed for interactive agent logic

FAQ

Q: Is ZeroLang Turing-complete? A: Yes, it supports loops, conditionals, and recursion while adding agent-specific constructs on top.

Q: Can it call existing APIs and tools? A: Yes, tool registries let ZeroLang programs invoke any HTTP API, CLI tool, or MCP server.

Q: What is the performance like compared to Python? A: The C-based VM executes significantly faster than interpreted Python for orchestration logic.

Q: Is it production-ready? A: ZeroLang is under active development by Vercel Labs. The core language and runtime are usable, with the ecosystem still growing.

Sources

• https://github.com/vercel-labs/zerolang
• https://github.com/vercel-labs/zerolang/blob/main/README.md