Building Always-On Agents (claude.com)

A new wave of always-on agents has emerged, showcasing their stateful and proactive capabilities that allow them to manage user interactions without direct supervision. Detracting from the notion that they are continuously operational, these agents spend most of their time inactive but wake when triggered by events from various sources. The significance of this resurgence lies primarily in recent advancements in AI models, which enhance judgment and reduce operational costs for long-term tasks. This has prompted increased interest from companies looking to adopt these proactive agents in their products, whether by upgrading existing systems or starting anew. The architecture of always-on agents revolves around five core primitives: event inbox, timers, memory, permissions, and outbox. These elements facilitate task execution, continuity across sessions, and manage user interactions with a proactive approach. For instance, the event inbox enables the agent to receive diverse triggers while maintaining a single active worker to ensure task integrity. Timers help with follow-ups or scheduled tasks, while memory stores essential user preferences and ongoing task progress. Additionally, permissions ensure safe execution of actions through user input verification. This holistic framework allows for more efficient, context-aware AI agents that can revolutionize user experiences in a variety of applications.