The era of agents building programming languages is exciting.
What I’m working on is a language agents can write, fix, and grow code with over time.
Almide started as an experiment, but it is becoming a practical compiler.
10 weeks. 2,500 commits. Two targets: native and WASM.
Almide is designed around one core metric:
Modification Survival Rate.
When an LLM changes existing code, how often does the result still compile, run, and pass checks without human fixes?
That metric shapes the whole system:
— predictable syntax: pipes, tuple pattern matching, expression-based design
— LLM-oriented diagnostics: repair suggestions the model can apply directly
— hallucination-resistant stdlib: 381 functions and aliases for common mistakes
— type-level effect tracking: pure vs effectful, enforced by the compiler
— context-friendly documentation: a one-file cheatsheet that fits in a prompt
— native and WASM from the same source
A GPU UI framework, Ceangal, already runs on the WASM target.
This is not just theoretical. It already runs.
The goal is not only to show that agents can build languages.
The goal is to build a language agents can use correctly over time.
If we are serious about languages for agents, I think we should measure how accurately LLMs can write and modify code in them.
Almide is my attempt to take the LLM-native programming language space seriously.
github.com/almide/almide
#Almide #WASM #ProgrammingLanguage