Vercel Labs has open-sourced the experimental programming language Zero on GitHub, which has received 5,310 stars and 339 forks. However, its real selling point is its positioning — this is a programming language "for AI Agents, not humans." Traditional programming starts with text, while Zero changes the source to a graph: Agents query the program graph, submit validated patches, and the compiler either accepts or rejects them. Humans only need to occasionally look at the text projection.

image.png

The compiler is in the loop, and humans are outside the loop

The core of Zero is to turn an Agent's editing into a patch operation on a semantic graph, rather than modifying lines of text. Each patch comes with a graph hash and expected field values, and the compiler verifies before writing: whether the corresponding node is still in the state it was when you read it, and whether the type and shape are correct. If the graph has been altered elsewhere or the expectations don't match, the edit fails immediately and is never written to disk. This gives the Agent's edits semantic anchors — the goal is to target "the value field of expression 653eeb6e," not "the third character on line 47," which is much more reliable than line numbers.

Text is now a secondary role: the .0 file is just a projection of zero.graph. Humans use export to view it, while Agents use query to read the graph and patch to write to it. import/export is simply the interface between humans and Agents. The community's reaction on Hacker News has been quiet, with doubts focusing on two points: why should an Agent learn a new language instead of using familiar ones like Python or TypeScript, and the advantages of Zero's semantic graph lack quantitative proof. As of its release, it has only 13 releases, and the language is still rapidly evolving, making it more of a thought experiment — this language will only be considered solid when it proves that "writing with Zero has fewer errors and uses fewer tokens than Python."