It started when taking the train to work and regularly hitting internet dead zones, so I initially downloaded Ollama and OpenCode to see if I could keep building while offline.
The first iteration was an installer that worked out what models I could reasonably run locally (as I wasn’t really sure what my machine would be able to do).
After that, I kept running into cases where an agent would say it had completed a task with a detailed description of the changes it had made but the file hadn’t actually changed.
The project then expanded into an Orchestrator inside OpenCode. It decides how complex a task is and delegates parts of it to subagents, potentially using different local models. Depending on the task, agents can then explore, plan or research something before passing to a coding agent.
The coding agent updates then gets passed to a verifier and complex changes can also go through a separate reviewer.
So, for simple task:
Coder -> Verifier
or for something more complex:
Explorer -> Planner -> Coder -> Verifier -> Reviewer
There’s also a Researcher when more knowledge is needed.
It’s still an experiment as I’m not sure yet how far this approach can push smaller local models.
some benchmarks would be nice.
also consider adding some even smaller models too: minicpm, lfm
I'll definitely look at adding both of those!