Vercel's "Zero" agent-first language is just Zig with extra steps

2026-05-20 Wed ⏱ 13 min @ThePrimeTimeagen

https://www.youtube.com/watch?v=RnxTTF8x-24
Original title: Why Did They Build This?

Prime reacts to Vercel's just-released "Zero", a new systems language pitched as "the programming language for agents". After reading the marketing copy, skimming the compiler, and writing hello-world, he concludes the design goals describe Go almost word-for-word while the syntax is Zig with renamed keywords — and nothing in it is actually agent-first.

What Zero claims to be

Agent-first learnability, small regular surface, deep stdlib.
"Deterministic tooling" (Prime: never heard of nondeterministic tooling).
Explicit capabilities, JSON diagnostics, type-safe fixes.
One-obvious-way regularity — Prime: that's Go.

What it actually looks like

Syntax cribbed from Zig: check ≈ Zig's try, raises as function colour, ~mem~-style comparisons.
void is a type, not a keyword — Prime "triggered AF".
Verbose examples → bigger token footprint, not smaller. Pollutes context windows faster.
No syntax highlighting on GitHub yet (too new).

Why "agent-first" doesn't hold up

Zero training-data presence — LLMs have never seen it. Worse for agents, not better.
Verbose surface = more tokens per program.
JSON diagnostics aren't a magic unlock; LLMs already parse C++ template vomit fine.
TypeScript wins because of a billion examples; Rust wins because the borrow checker constrains output. Zero offers neither.

Compiler red flags

~15 nearly-identical vector implementations (push logic duplicated across types).
Backends (a64, elf, etc.) repeat near-identical code with subtle naming drift.
Diag handling inconsistent across backends (one checks pointer existence, another assumes always-exists).

Prime's take