The author identified two primary causes for system performance degradation while running AI agents like Claude Code and Codex. The Windows Defender antivirus was identified as a major factor, as it scans every file created or modified by the agents. Given that agents generate tens of thousands of files during tasks like package installation and project builds, the Antimalware Service Executable process consumes significant CPU resources.
To mitigate this, the author excluded specific working directories and agent programs from Defender's real-time scanning. Additionally, the author utilized the Dev Drive feature on Windows 11 for a more efficient scanning mode. These adjustments reduced CPU utilization from 67% to approximately 12% without requiring hardware upgrades.
The second issue involved storage overflow on the system drive C. Although agents were configured to operate on drive D, their associated tools and processes frequently wrote cache and temporary data to the AppData folder on drive C. The author resolved this by cleaning npm and Docker caches and moving these directories to drive D using symbolic links.
The author performed diagnostics using the Local Commander tool, though noted that manual checks via PowerShell are equally effective. The article also mentions the release of open-source projects GPU-unlock and GPU-unlock-ADA, which remove manufacturer restrictions on matrix calculations for Nvidia RTX30xx and RTX40xx series graphics cards.
