Zero (zerolang)

An experimental language from Vercel Labs — the research arm of the company ../cloud-wiki already pages as a frontend PaaS and the maker of Next.js — built on one refusal: the program is not text. The authoritative representation is a semantic graph the compiler owns (zero.graph); .0 files are projections of it, generated so a human can read a change, and explicitly not the interface the language is designed for. The tagline on the site is “an agent-first language experiment”, and the front page and the repository agree on what that means — the intended writer is an LLM agent, and the text file is a courtesy extended to the reviewer.

The mechanism

An edit is a checked patch, not a save. zero patch targets semantic nodes by hash and expected value; a patch built against a stale view of the graph is rejected before anything is written. Reading is symmetrical: zero query returns symbols, types, effects, call edges and ownership facts from the graph directly, rather than requiring the writer to re-derive them from source text. The rest of the CLI is check, test, run, inspect, and import/export for crossing the projection boundary in either direction.

The claimed payoff is stated as a loop property rather than a language property: an agent otherwise writes text, runs separate tools to learn what it broke, and guesses in between. Here the compiler sits inside the edit, so the failure surfaces at patch time. Vercel’s phrasing — “reduce agent guessing without relaxing runtime goals” — pairs that with a conventional performance list: token efficiency, low memory, fast startup, fast builds, low latency, no dependencies.

What it actually is today

Apache-2.0, ~5.2k stars, ~338 forks, 49 open issues (2026-08-06). The repository’s own warning is unusually direct — experimental, breaking changes, rough edges, security issues, run it in an isolated workspace and not against production. The worked examples are at the level of initialising a project, adding a function, printing output, and running the result. So the design is legible and the language is not yet evidence that the design works: nothing here demonstrates a program large enough for graph-first editing to be either an advantage or a burden.

Why it matters here

Every other source in this corpus takes text as the substrate and argues about what the text means — memory-management-model, concurrency-model, hosted-language, language-governance are all questions asked of a textual language. Zero answers a question none of them asks: what is the program stored as, and who is allowed to edit it directly? (programming-language).

It has one relative in the corpus. clojure‘s code-as-data makes the program a data structure the program itself can manipulate at compile time; Zero makes it a data structure the toolchain owns and the writer reaches through an API. Both drop the premise that source text is the real artefact, and they do it for opposite beneficiaries — the macro author versus the compiler. Zero is also, so far, the only source here whose target audience is not a human programmer.

Tier caveat. T1 as first-party (site plus the project’s own repository), and the standing rule applies with force: authoritative on design intent — the graph model, the checked-patch contract, who this is for — and unmeasured on everything else. “Token efficiency”, “fast builds” and the argument that agents guess less carry no benchmark, no comparison and no study of agents using it. Star count and issue count are adoption signals about attention, not use.

Related

programming-language · clojure · language-governance · synthesis · cross-spoke: ../cloud-wiki holds the canonical Vercel page (as a hosting provider), ../agentic-tooling-wiki owns the harnesses that would have to adopt this, ../dev-tooling-wiki owns compilers as build infrastructure

Governance footnote

Vercel Labs is a fifth arrangement next to the four in language-governance: not a standards committee (c), a foundation (rust, nodejs), a public RFC process (php) or a company funding a core team (go, clojure, elixir), but a company research org with no stated process at all — because there is no constituency yet to have one for. Whether that becomes a model or stays a phase is the succession question language-governance says nothing in the corpus answers.