Vercel's "Zero" Is a Programming Language Built for AI Agents

Vercel Labs released Zero, an experimental systems language designed from the ground up for AI agents. The premise is simple: every mainstream programming language was built for humans, so why are we making bots write Python and TypeScript? Zero flips that assumption by giving agents a language where the compiler output is structured for machines to parse and act on.

What Zero Actually Does

The GitHub repo at vercel-labs/zero describes Zero as "a systems language for small native tools, explicit effects, predictable memory, and structured compiler output." Three design choices stand out from the project's site at zerolang.ai:

JSON-native diagnostics. When Zero's compiler hits an error, it outputs structured JSON instead of human-readable console messages. An agent can parse that output directly, figure out what went wrong, and fix its own code without scraping error text.

Explicit everything. No hidden allocator, no implicit async, no magic globals. If a function touches the outside world, the signature says so. The language uses capability-based I/O where functions declare what they touch and the compiler rejects unavailable capabilities at compile time. This makes the code predictable for an agent that needs to reason about what a program will do.

Small native artifacts. Zero uses static dispatch, explicit capabilities, and no mandatory garbage collector or event loop. It is built for tiny tools, not large applications. Size reports make artifact costs visible.

Still Experimental

The repo README is direct about the project's maturity: "Zero is experimental and still changing. The compiler, standard library, docs, and examples are useful for trying the language and giving feedback, but the language is not stable yet." At the time of the video, the repo had 941 stars and 51 forks under an Apache 2.0 license.

Key Takeaways

• Vercel Labs launched Zero, a systems language where the compiler produces structured JSON output that AI agents can parse and act on directly
• The language enforces explicit memory, effects, and I/O capabilities so agents can reason about programs without hidden behavior
• Zero is experimental and unstable, built for small native tools rather than full applications
• The broader question the video raises is worth sitting with: if agents are writing most of the code, should the languages they use be optimized for them instead of for us?