AgentStop is a lightweight efficiency supervisor developed by Brave that monitors local LLM-based agents in real time and terminates unpromising runs early to reduce battery drain on consumer devices. It works by tracking token log-probabilities, token counts per step, and token overlap between steps — signals already produced during normal inference — and feeds them into a gradient-boosted XGBoost classifier. Evaluated on web question answering (FRAMES, SimpleQA) and coding (SWE-Bench Verified), AgentStop reduces wasted energy by 19–23% with less than a 5% drop in task utility. The supervisor itself costs under 0.01 mWh per inference call. The project is open source and was accepted at ACM CAIS 2026.
Table of contents
SummaryLocal AI Agents are Necessary for PrivacyBut Local AI Agents Are Expensive to RunAgentStop: Brave’s First Step Towards Building An Efficiency Supervisor for AgentsAgentStop reduces energy consumption at minimal utility cost211 Impressions