AI & agents
Anthropic shares a design guide for always-on agents
- Published
- Source
- Anthropic (claude.com)
Summary
Anthropic published a guide for building always-on agents, which stay dormant most of the time and wake up when work arrives. The architecture combines five primitives: an event inbox, timers, memory, permissions, and an outbox. The company says better model judgment and falling costs have made the pattern practical, and cites one of its own long-horizon coding benchmarks in which Opus 5.5 cost 78% less per solved task than Opus 5.
Why it matters for our work
Agents that wake themselves and keep working could change how daily work gets done, beyond on-demand assistance. The design advice and cost figures are Anthropic's own, but the guide is directly usable for teams evaluating always-on agents.
Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.