More inference compute does not reliably rescue local computer-use agents
Researchers tested longer context, longer trajectories, task decomposition and parallel attempts for selected Qwen3-VL, UI-TARS and OpenCUA models on OSWorld. Extra context saturated, long runs prolonged some wrong paths, decomposition added overhead and parallel attempts were expensive.
Why it matters: More steps and context can spend additional compute on the same mistaken plan. Reliability needs failure detection and selective escalation, not an unrestricted execution budget.