A programming language built for LLM agents, not humans
zerolang treats source code as durable data but gives agents a compiler-derived graph to inspect and edit, so they stop guessing with text patches.

What it does
zerolang is an experimental systems language with a split personality: humans read and review .0 source files, while agents interact with a compiler-derived “ProgramGraph” of node IDs, types, effects, and hashes. The CLI exposes structured inspection and checked edit commands—zero graph dump, zero graph patch, zero check --json—so agents can target semantic nodes instead of praying their line-range patches didn’t break ownership or imports.
The interesting bit
The graph-edit loop is the real departure. An agent submits an operation against node #expr_653eeb6e with an expected graph hash and field value; the compiler validates, lowers, writes, formats, reparses, and rechecks before touching source. It’s a bit like database transactions for code changes—optimistic concurrency control against stale context.
Key highlights
- Source remains the stored truth; ProgramGraph is derived interchange data, not a separate file format
- Graph patches target semantic nodes with preconditions (hash + expected field value), not raw text ranges
- CLI ships version-matched “skills” (language rules, diagnostics, stdlib) bundled with the compiler binary
- Structured JSON output for tokens, parse, check, graph, size, and repair-plan commands
- Explicit design goals: token efficiency, low memory, fast startup/builds, zero dependencies
- Experimental and intentionally unstable; breaking changes preferred over compatibility layers
Caveats
- Security vulnerabilities “should be expected”; explicitly not ready for production or sensitive data
- Source rewriting after graph patches is only supported for “canonical
.0source”—unclear how much of the language that covers - The README shows two different graph hashes and node IDs for the same example, suggesting the output isn’t stable across versions (or the docs drifted)
Verdict
Worth watching if you’re building agentic coding tools and tired of watching LLMs flail with regex refactors. Skip it if you want a language to actually ship product with—this is research infrastructure with a C compiler and a warning label.