agentlanguages.dev
verification camp · also syntactic

Zero.

Vercel Labs' agent-first systems language. Sub-10 KiB native binaries. Structured JSON diagnostics with stable codes and typed repair plans. One obvious path.

authorChris Tate and Matt Van Horn / Vercel Labs
implementationC (zero-c bootstrap); self-hosted compiler-zero in progress
targetNative binaries (direct ELF/Mach-O/PE emitters, no LLVM), WebAssembly
licenceApache-2.0
first seenMay 2026
maturityearly implementation

The thesis.

Zero is Vercel Labs’ bet that the bottleneck in agentic coding is not the language but the channel between the compiler and the agent. The standard loop is fragile: the compiler emits prose written for human engineers, the agent parses it as text, the agent guesses at a fix, the next compile produces a new prose error in a slightly different format. Zero’s answer is to replace the prose channel with a structured one. zero check --json emits a stable error code (NAM003), a human-readable message, a line number, and a typed repair object an agent can act on. zero fix --plan --json returns a machine-readable edit plan. zero explain returns structured explanations against the installed compiler version.

Humans read the message. Agents read the JSON.

The distinctive move sits at the language level, not the toolchain level: Zero collapses the syntactic and verification camps into a single product decision. The language documents itself as preferring “one obvious way to express most things, even when that makes code more explicit than a human might choose,” which is syntactic-camp framing; but the obviousness is bought with capability objects on main, explicit raises markers, and effect-visible signatures, which is verification-camp machinery. Where MoonBit, the catalogue’s other industrially backed verification entry, invests in semantic-aware token sampling, Zero invests in making the surface area small enough that an agent doesn’t need help generating it in the first place.

What it looks like.

pub fun main(world: World) -> Void raises {
  check world.out.write("hello from zero\n")
}

There is no hidden global process object. world: World is an explicit capability passed in by the runtime; raises declares the function can propagate errors; check handles a fallible operation. A function that doesn't ask for World cannot write to stdout.

Distinctive moves.

Maturity.

v0.1.1, Apache-2.0, released 15 May 2026, 3.3k stars on vercel-labs/zerolang at first cataloguing. The README and homepage are explicit that this is a “pre-1 experiment”: syntax and APIs are not a contract, breaking changes are expected, and security vulnerabilities should be expected — Vercel Labs recommends running Zero only in isolated environments. The repo maintains two compilers: zero-c, the C bootstrap; and compiler-zero, a self-hosting compiler written in Zero. Cross-compilation is limited to a documented target subset; there is no package registry yet; VS Code syntax highlighting ships in-repo. Named contributors are Chris Tate and Matt Van Horn.

The bet is that structured agent-first compiler output becomes table stakes once developers see what it does for repair loops. Even if Zero itself doesn’t win, the design pattern — stable codes, typed repairs, version-matched skills — is a concrete argument for what every other compiler should ship.

Agent tooling.

The toolchain is the agent tooling. zero check --json returns diagnostics; zero explain <code> returns explanations; zero fix --plan --json returns edit plans; zero skills get zero --full returns version-matched workflows. zero graph --json, zero size --json, zero routes --json, and zero doctor --json round out the inspection surface. Vercel’s separately released skills.sh ecosystem is consumable by Claude Code, Cursor, Codex, and other agent harnesses through the same Agent Skills spec that zero skills follows.

design DNA
  • MoonBit verification Industrial backing parallel. Vercel Labs and IDEA Shenzhen are the two best-resourced bets in the catalogue; MoonBit invests in semantic-aware sampling, Zero invests in structured compiler output and version-matched skills.
  • NERD syntactic Cross-camp foil. Both lean on a small keyword vocabulary and 'one obvious way' framing; NERD does it for syntactic legibility, Zero does it inside a verification project with capability-typed effects and a typed repair API.
  • Boruna orchestration Structured-diagnostics parallel. Zero ships JSON diagnostics with typed repair IDs at the language level; Boruna ships hash-chained evidence bundles at the runtime level. Both reject prose as an interface for agents.