Stanford CS146S: How always-on agents work

Always-on, proactive, and persistent AI agents represent a new era of development where models operate ambiently in the background on dedicated cloud machines with their own files, memory, and computer-use capabilities. Unlike terminal-based harnesses that require constant human prompting and oversight, these agents function more like colleagues that can handle long-running tasks, coordinate sub-agents, and execute multi-step workflows over days or months. Advancements in instruction following, computer use, and auto-review safety checks have enabled these systems to act autonomously while maintaining security. Open article →