Zerolang
Zerolang is an experimental Vercel Labs language meant for agents. The semantic graph is treated as the program itself, and .0 files are a human-readable projection. Agents are supposed to patch graph nodes instead of fighting line-based text diffs. The pitch includes native binaries and agent-friendly diagnostics. Host backend quality is still uneven.
Why it is on this list
It sits next to scriptc and Native in the same Labs wave. Marketing says agents can safely edit a semantic graph. On the issue tracker, people have said the code quality is low and you should not run it on your own machine (#181). Maintainers agreed and said they were hiring. Later reports list concrete bugs: big integer math that comes out wrong, stack frames that get cut short and crash, and safety checks that disappear. Long write-ups in Chinese add that whole compiler trees were deleted after criticism, and the runtime is thin. Marketing says stale hashes are rejected before write; #425 is about wrong module hash comparison. Those two sit poorly together.
References (10)
vercel-labs/zerolang repository
~5.2k stars, 100+ open issues; homepage zerolang.ai; "Programming Language for Agents".
README safety warning
Official warning — experimental, expect breaking changes and security issues; isolate, not production/sensitive data.
Low code quality. Don't run this on your machine.
Community quality blast (contracts, C pointer effects, stringly IR); maintainer acknowledged and noted hiring.
Bug audit — 18 reproduced findings
Single audit of tokenizer/parser/borrow/backends/stdlib: S1 wrong-code i64 CMP on darwin-arm64; S2 silent stack-frame truncate → SIGSEGV; S3 borrow soundness on dynamic array set; plus parser/stdlib wrong results.
darwin-arm64 Mach-O direct backend missing ops (CGEN004)
Many direct-* examples fail on Apple Silicon host backend — missing byte ops, error/raise paths, etc.
Incorrect module hash comparison
Graph/module hash comparison bug — core to the "stale hash reject" agent story.
Major syntax changes in 0.1.4
Early versions still thrashing syntax — agent skills and human projections both chase a moving target.
Vercel Labs marketing framing (agent graph, not text)
Official X account pitch — graph-first, verified refactor; experimental framing often buried under launch energy.
PL Nerd 迎来惊天大变 · Vibe Coder 喜提 AI 原生（zerolang 扒代码）
Chinese long-form roast of Zero/zerolang as an "agent-native" language whose implementation reads like anti-hype performance art — documents crude IR and compiler structure that undercut the graph-first marketing.
如何评价 Vercel 近期发布的 Zero 编程语言？（续扒）
Follow-up: compiler-zero tree deleted after critique; time/rand/fs still modeled in IR; new "runtime" is mostly http/json while hardcoding remains; giant EmitContext/RuntimePatch surface; ownership/effect-handler fashion on a thin checker — community evidence that agents thrash the tree faster than quality lands.