2) I self host a Q4 quant of GLM 5.3 flash, at work using Antigravity and Gemini Argon

3) Vanilla Pi cli and running in a VM. I customized it to work around common LLMism, like intercepting and modifying requests or tool calls.

4) Pi is my agent orchestration, it can run in headless mode recursively. Also most of the things the agent UIs provide can be done via writing to shared files to disk

5) No