LEASH introduces a training‑free framework for weak‑to‑strong escalation and oversight of computer‑use agents. It operates in three stages: first, an initial task‑level uncertainty score routes the request to either a weak or strong model; second, step‑level uncertainty computed during the weak agent’s action generation triggers runtime handoff to the strong model when the latent state resembles historically failed trajectories; third, after escalation a shadow observer monitors the black‑box strong agent’s continued behavior using the same uncertainty signal. The uncertainty estimator reuses the hidden activations already produced for action selection, requiring no extra forward pass or dedicated monitor training. Evaluated on the OSWorld benchmark with four contemporary model backbones, LEASH outperforms a standalone weak agent and exceeds random escalation by 6.55–11.94 success points at equivalent escalation frequencies. Using Qwen3.5‑35B‑A3B as the strong model, LEASH attains 73.66 % task success at a cost of $0.4117 per task, matching the performance of an always‑on Claude Opus 5 (72.90 %) while cutting inference expenses by 19.2 %. These results demonstrate that latent uncertainty can drive cost‑effective escalation and sustained oversight without additional training overhead.