Summary
- Always-on agents are stateful, proactive AI systems that manage context across sessions and execute tasks without direct supervision.
- Their architecture relies on five core primitives: an event inbox, timers, memory, permissions, and an outbox.
- Recent model advancements and lower inference costs have made running always-on agents more practical for long-horizon tasks.
- Runs operate while the agent wakes up periodically from sleep, processing events and storing important context for continuity.
Why it matters
Understanding the architectural primitives of always-on agents enables developers and engineers to build proactive, stateful AI products that operate reliably across disjointed background runs.