Zero

A tiernew this week

A systems programming language built from the ground up so AI agents can read compiler output, repair errors, and ship native programs — no human translation required.

Open Zero →Compare with alternatives

Kai's verdict

A genuinely interesting bet on what programming languages look like when AI agents are the primary consumer of compiler output — but it's v0.1 experimental territory, so treat it as a thought-provoking sandbox, not a production dependency. (Verdict pending Phi's full review.)

Strengths

Compiler outputs structured JSON diagnostics with stable error codes by default — not as an opt-in — so agents can act on errors deterministically without LLM inference over prose
Capability-based I/O makes every side effect explicit in function signatures, letting agents understand what code does without tracing call graphs at runtime
Single unified CLI binary with consistent --json flags across all subcommands (check, build, graph, size, doctor, skills) — predictable for agentic toolchains
Compiles to sub-10 KiB native binaries without LLVM, no GC, no hidden allocator — genuinely lean output
Built-in `zero skills` command outputs agent-consumable capability descriptions compatible with Claude Code, Cursor, Codex, and 17+ AI coding assistants

Weaknesses

Deeply experimental (v0.1.x, language not stable) — no package registry, no stable compiler spec, limited cross-compilation support
Memory safety is a design goal but implementation is early-stage; borrow checker is described as basic compared to Rust-grade guarantees
The core premise is debatable: modern LLMs already parse Rust/Go errors well enough, so the actual agent reliability gains over existing languages remain unproven

Best for

AI engineers and agentic coding tool builders who want to experiment with a language where the entire toolchain — diagnostics, repair hints, capability declarations — is designed as machine-readable data for agent consumption.