Anthropic outlines architecture primitives for building always-on AI agents

• Anthropic identifies five core primitives—event inbox, timers, memory, permissions, and outbox—that enable agents to operate proactively and maintain state across sessions.
• Recent model advancements have made always-on agents practical, with Opus 5.5 costing 78% less per solved task on long-horizon coding benchmarks than Opus 5.
• The system operates on a cycle where the agent sleeps most of the time, waking up for brief runs triggered by events or timers to perform work and update stored state.
• Event inboxes use durable lists and workers to handle triggers safely, preventing race conditions and ensuring events are not lost if the worker crashes.

Architecture Components

• Timers are stored separately from the inbox and managed by a scheduler that polls for due follow-ups or recurring schedules to trigger new runs.
• Memory is implemented as a directory of files mapping to durable storage, storing user instructions, notes, past messages, and task progress to maintain continuity.
• Permissions manage unattended risk through credential proxies, auto-mode classifiers for high-risk actions, and deferred tool calls that pause execution for user approval.
• Outboxes handle message delivery, allowing the agent to pace communications and recover from errors, with a recommendation to start quiet and earn user attention over time.