However, I wanted to give Zig a try in an agentic coding scenario. For tasks that would take a few seconds when choosing Python, Java, or JavaScript as a target language, it would take tens of minutes and waste millions of tokens before producing anything.
Almost any model gets stuck trying to figure out the correct syntax and correct libraries for a specific Zig version, fighting with compiling and figuring out function call parameters, frequently taking it wrong and going on side quests for things that should just work.
I guess the relative lack of resources and the language instability don't play well for models that try to generate Zig code. Using specific tools like zig-mcp helps only a bit.
Until LLM support for Zig improves (one needs to spend significant resources for that to happen), LLM-generated Zig code won't be good enough for either Zig programmers or Zig contributors.