Anthropic explains the architecture behind always-on agents

Anthropic details the five core primitives behind always-on agents: an event inbox, timers, memory, permissions, and an outbox. The post says recent model gains made these agents more practical, citing Opus 5.5 costing 78% less per solved task than Opus 5 on one long-horizon coding benchmark versus 40% less on typical workloads. It also covers risks in unattended permission handling, including credential proxies, auto mode classifiers, and deferred tool calls that pause a run for user approval.