LEASH Escalates to Frontier Models Only on Latent Uncertainty, Matching Always-On Claude Opus 5 on OSWorld at 19.2% Lower Inference Cost
A new arXiv paper keeps a small open-weight agent in control of computer-use tasks and hands off to a frontier model only when the weak model's hidden states resemble past failures. Authors report OSWorld success matching always-on Claude Opus 5 at 19.2% lower inference cost.