I'm currently working on a fully autonomous software development client. You can just tell the LLM what you want it to build, optionally adding a design doc or screenshots of app designs, and it iteratively keeps building the app and adding new features until you tell it to stop. It uses a vision model to verify correctness and rate the UI in-between iterations.
The best part is that it can run purely on relatively small local models (I ran it on Qwen2.5 Coder 14b and Qwen3 VL 8b using my MacBook).