Someone finally built a programming language for the code writer that actually matters now: the LLM. Vera, created by Alasdair Allan, throws out assumptions baked into every language humans use and asks a different question. What if the language was designed so models can't screw up?

The biggest win is killing variable names entirely. Instead of names, Vera uses typed De Bruijn indices like @Int.0 and @String.1. No names means no naming hallucinations, which the Vera team identifies as one of the most common failure modes for code-generating models. Every function also requires explicit contracts: preconditions, postconditions, and effect declarations. The compiler verifies these statically using Z3 where possible, with runtime fallback for things it can't prove at compile time.

The design philosophy is blunt. The model doesn't need to be right. It needs to be checkable without developers abdicating their oversight role. Fewer valid programs means fewer ways to be wrong. Every error message is written in natural language with a concrete fix, aimed at the LLM that wrote the broken code, not a human reading over its shoulder.

Vera compiles to WebAssembly and ships with built-in support for HTTP requests, JSON parsing, HTML scraping, and LLM inference calls to empower agents making external calls. Network I/O and model calls are typed algebraic effects that show up in function signatures, so a function calling an LLM has to declare that fact.

According to VeraBench, models hit up to 100% accuracy writing Vera compared to 86-96% in Python and TypeScript. Bold claim. Those are the Vera team's own numbers, run on their own benchmark suite, and we don't have details on what tasks were tested or how difficult they were. But the underlying logic holds. When you eliminate naming decisions and ban implicit behavior, models have fewer ways to fail.

The code looks alien to humans, but that's the point. Vera isn't competing with Python for developer experience. It's competing for reliability in generated code, and the early approach makes sense even if we need independent verification of those headline numbers.