Vercel Labs Releases Zero, a Graph-First Systems Language for AI Agents

Introduction of Zero Language

Vercel Labs has released Zero, an experimental systems programming language. The language is built on the premise that AI agents, rather than humans, will be the primary readers of compiler output. Chris Tate of Vercel introduced Zero on May 15, 2026, describing it as a systems language that is "faster, smaller, and easier for agents to use and repair."

Technical Specifications and Features

Zero uses the .0 file extension, is licensed under Apache 2.0, and compiles to native binaries for Linux, macOS, and Windows. Early reports indicate a small footprint, with a 'hello world' binary at 16.2 KiB and a millisecond build time. A core feature is its toolchain contract, where every subcommand of the 'zero' binary supports a '--json' flag and a unified diagnostic schema, providing stable error codes and typed repair metadata like 'declare-missing-symbol'. The 'zero fix --plan --json' command returns a machine-readable repair plan for agents.

Graph-First Workflow and Explicit Effects

Version 0.3.0 of Zero introduced a graph-first authoring workflow, making a binary 'zero.graph' store the primary compiler input. '.0' files serve as human-readable projections. Agents interact with the code via 'zero query' and 'zero patch', with patches secured by graph hashes to prevent stale or invalid edits. Additionally, Zero enforces explicit effects; any function interacting with external systems must accept a 'World' capability, which the compiler verifies, making network, filesystem, or standard output access visible in the function signature.

Development and Community Reception

The project has rapidly evolved, reaching v0.3.4 and accumulating over 5,200 stars on GitHub. Significant changes have occurred across versions, including the adoption of row syntax in v0.1.4, promotion of canonical '.0' text in v0.2.0, and the rejection of source projection inputs at the compiler boundary in v0.3.0. Existing text-first packages now require 'zero import' to convert source into the graph, with 'zero export' and 'zero verify-projection' for human review and CI drift checks. Version 0.3.2 improved 'zero import' speed by approximately 12x for large programs. Community discussions have noted the capabilities feature and debated the novelty of structured errors, with proponents emphasizing the agent-centric design.