HarnessMinusMinus-V: Recursive Self-Improvement for Visual Agent Harness Token Efficiency
Yuhua Jiang, Chang Xu, Feifei Gao, Li Zhao
Abstract
Visual agent harnesses repeatedly invoke models to reprocess context and regenerate resolved actions, increasing token consumption. We introduce HarnessMinusMinus-V (HarnessMM-V), a recursive self-improvement (RSI) framework that reduces visual harness token cost. Training feedback improves the construction and execution of Executable Decision Graphs (EDGs), which store resolved actions with visual checks and dependencies. Guarded Decision Reuse (GDR) retains graphs, checks current pixels and confirms completed predecessors; Certified Callback Substitution (CCS) executes eligible actions without another model call. Native scoring measures success; separate validation selects the harness before testing. On a predeclared 546-task VisualWebArena test partition, execution cost falls from 74.31 to 55.27 Tokens/task (k), with 15.93\% versus 14.84\% success. Controlled visual ablations isolate the effects of graph persistence and model-call reduction. Transfer to 30 non-form website tasks under five visual conditions saves 14.41\% in execution cost. Code is available at https://github.com/YuhuaJiang2002/HarnessMM.