Brave has developed 'AgentStop,' a technology that solves the problem of local AI running indefinitely and unnecessarily consuming battery and GPU resources.
While cloud-based services like ChatGPT and Claude tend to attract attention in AI-related discussions, there is also a lot of activity in developing local AI that can run without leaking information to external parties. The developers of the privacy-focused browser 'Brave' are also working on local AI development, and on May 28, 2026, they announced the development of ' AgentStop ,' a system that interrupts the unnecessary execution of local AI agents to reduce battery consumption.
AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices | Brave
When running AI in a local environment, it naturally consumes your PC's computing resources and battery. For simple chat applications, this isn't a major problem because processing stops once the response is output. However, when running it as an AI agent to perform complex tasks, it requires repeatedly executing long processing steps, and patterns where 'the process ultimately fails after many repeated inferences' occur frequently.
The graph below shows the system load progression when 'Qwen3-Coder-30B-A3B' is run as an AI agent on a MacBook Pro equipped with an M1 Max processor. It reveals that 'processing continues for more than 10 minutes,' 'more than 30 LLM inference calls are made,' 'GPU power consumption sometimes exceeds 40W,' and 'GPU temperature remains above 90 degrees Celsius for extended periods.' In other words, the local AI agent occupies the MacBook Pro's CPU and GPU for long periods, consuming a significant amount of battery power. Often, even after running long processing times, the task ultimately fails, resulting in wasted computing resources and battery consumption.
Therefore, the Brave development team developed 'AgentStop,' a system that detects task failures in local AI agents early and interrupts their execution. When a local AI agent fails a task, 'signs of failure' appear, such as 'low confidence in output tokens,' 'an abnormally high number of processing tokens per step, leading to a loop,' or 'repeated output of the same result.' AgentStop monitors the output of the local AI agent and interrupts its operation when signs of failure appear, thereby reducing battery consumption and other factors.
When using Qwen3-Coder-30B-A3B as a local AI agent to run 500 tasks in the benchmark test 'SWE-Bench Verified,' enabling AgentStop reduced power consumption by approximately 19% compared to when it was disabled, and the decrease in task completion rate was limited to approximately 3%.
Brave positions AgentStop as 'the first step towards making local AI agents consider not only privacy and convenience but also energy efficiency.'
AgentStop is being developed as an open-source project, and its source code is available at the following link. It is licensed under the MIT License.
GitHub - brave-experiments/AgentStop: AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices · GitHub
https://github.com/brave-experiments/AgentStop
Related Posts:
in AI, Posted by log1o_hf