Current:
Option A)
Unsloth Studio allows you to connect your local model, like Qwen, to an already popular harness like Claude Code and Codex. Then tell the model to do something like 'find the fastest way to do X, don't give up' and it will loop through this process many times until it reaches an end point after trying 300 different things.

Option B)
Using Unsloth Studio, tell the model to do something like 'find the fastest way to do X, don't give up' and then it will try 3-5 things and tell you it did it's best.

Future:
Do this entirely within Unsloth. Be able to load a model and have it run a local inference hardness engine for automated, self-correcting executions so that it continues to run things until it actually completes them.