I built an agent that builds agents — runs on local Qwen3, pure Python.

Hi, I built this agentic AI system.

Closed-loop — takes a raw idea, ships a standalone .py agent file.

What's different:

• Interviews you until it understands the request before building anything
• Two testing stages: prompt validation via LLM invoke, then real subprocess execution of generated code. Not the same thing.
• Self-referential: injects its own source as a reference template for generation
• Structured rating schema drives iteration. Human approval gate before anything saves.

Runs on Qwen3.6-35B A3B Q8_0 locally via llama.cpp.

GITHUB PAGE

Give it a shot and tell me what you think.