Hardware: I have a custom setup with a bunch of amd graphics cards. I will do a writeup about that sometime this year.

I use opencode exclusively. It’s probably not the best harness but it’s the one that gives users the most control and I like that.

The actual models I used for this work are Qwen3.8 27b - FP8 and Qwen3.8 - Flash - Next UD_Q6_K.

I wrote my own agents with duty separation. Flash is the orchestrator and long-horizon guide. While 27b is the faster, more expandable worker. This combination feels about as smart as an opus 4.8 with maybe not as great token efficiency. For simple tasks, i use the worker directly. I have a ton of repo documentation and my agents have access to read and write to Forgejo, both for git and api (issues and PRs respectively)

Maybe I need to write an article about this…