Skip to main content
Back

Vercel Labs Ships Zero, an Open-Source Language Built for AI Agents

Vercel Labs open-sources Zero, a graph-first systems language where the compiler speaks JSON so AI agents can write, check, and repair code without human translation

Repo, Full Access
  • What it is: Zero is an experimental systems language from Vercel Labs where the program is a semantic graph, not text files, making it natively editable by AI agents.
  • Graph-first design: Agents submit checked graph patches targeting semantic node IDs rather than line numbers; invalid or stale edits are rejected before they touch the store.
  • JSON compiler output: All diagnostics include stable error codes, typed repair metadata, and machine-readable fix plans — no prose parsing required by agents.
  • Systems-level runtime: Compiles to sub-10 KiB native binaries with no GC, no hidden allocator, no implicit async, and zero external dependencies.
  • Not production-ready: Explicitly pre-1.0 with breaking changes expected; not suitable for sensitive data or trusted infrastructure.
  • Free and open-source: Available now under Apache 2.0 at github.com/vercel-labs/zerolang, installable via a single curl command.

Every time a coding agent like Claude Code or Cursor hits a compile error today, it has to do something deeply awkward: read a message written for a human, infer what went wrong, and guess at a fix. When an AI agent writes code and hits a compile error, it has to parse unstructured text, infer what failed, and guess at a fix. That chain is fragile at every step. Vercel Labs just shipped an answer to that problem in the form of a brand-new programming language.

Vercel Labs released Zero , a low-level systems programming language whose compiler was built from the ground up to communicate with AI agents rather than human engineers. The project, built by Vercel engineer Chris Tate, asks a simple question: what does a programming language look like when machines need to use it as much as people do?

The graph is the program

The deepest idea in Zero is not its syntax or its small binary size. It is a fundamental rethinking of what a "program" even is. In every language you have used before, text files are the source of truth. Zero flips that.

In Zero, the graph is the program database. The graph stores declarations, types, calls, blocks, imports, capabilities, and source-map facts directly. Agents edit those facts with checked graph patches. Humans read .0 projections when they want a source-like review view. Think of it as the difference between editing a database record versus editing the HTML that renders it.

The graph gives agents explicit handles: symbols, node IDs, graph hashes, types, effects, ownership facts, capabilities, imports, call edges, and target facts. Edits can target semantic structure instead of line ranges. Instead of asking an agent to find line 42 and rewrite a string literal, you tell it to patch the literal node with ID #expr_653eeb6e. The compiler either accepts the patch or rejects it before anything is written to disk.

What the agent loop actually looks like

The traditional agentic coding loop is: write text, run