Plover makes GUI agent plans externally editable artifacts — finding that many autonomous failures are 'structurally repairable'

arXiv·medium signal

Submitted July 16, Plover uses a planner-executor split that externalizes task plans and replanning as persistent, inspectable, revisable artifacts rather than keeping them internal to the agent. Users supervise execution and make localized corrections through editable plans, combining natural-language guidance with screenshot-grounded intervention. The core claim is the interesting one for anyone building browser or computer-use agents: a large share of GUI agent failures are recoverable if the plan stays visible and intervention can be scoped narrowly instead of restarting the run.