Computer-use agents that handle long-horizon tasks face a cost-performance trade-off. Running a frontier model at every step is expensive, while a small open-weight model alone often fails on complex or extended work. This leaves developers choosing between high cost and low reliability.
The LEASH framework proposes a weak-to-strong escalation strategy. A weak model handles most steps, and a stronger model is called in only when needed. The key innovation is uncertainty awareness: the system monitors the weak model's confidence and escalates when it is uncertain, rather than doing so blindly or never.
This approach aims to provide a practical middle ground, reducing reliance on expensive frontier models while still tapping into their capability when necessary. The abstract describes the framework's motivation and design, though it does not include experimental results or comparisons with other methods.