The real insight over the past 2-3 days has been a combination of: 1. Local models like Qwen 3.6 and Deepseek v4 Flash (even in a 2-bit quant!) are quite useful, and understand integrations like tool calls well. 2. A harness like OpenCode is an immense step forward in keeping local models on-task. This site was crafted almost entirely locally with Qwen 3.6 35B-A3B. 3. It's becoming clear that adding on additional skills and integrations will pay huge dividends in getting it to operate autonomously for hours. I think a project like my good friend's https://dex.rip for cross-session memory and tracking will take this even farther.