I think a lot of SOTA models are already going this way. Long autonomous tasks in claude code / codex already do this to stay on track and avoid multiplying err

I think a lot of SOTA models are already going this way. Long autonomous tasks in claude code / codex already do this to stay on track and avoid multiplying errors over many steps.