Anthropic outlines primitives for always-on agents

The industry obsession with complex multi-agent swarms missed the obvious: production-grade autonomy requires boring, battle-tested distributed systems patterns rather than elaborate prompt chains. Anthropic formalizes how to build reliable persistent agents by treating the LLM as an intermittent worker rather than an infinite loop.

• –Sleep-by-default architecture: Framing always-on agents as dormant processes woken by an event inbox or scheduler eliminates runaway token burn and race conditions from simultaneous triggers.
• –Durable file-based memory: Mirroring Claude Code's auto-memory pattern, separating state into dedicated stores for user notes and task progress keeps context overhead lean and cache-friendly.
• –Deferred tool calls: Suspending sessions mid-call for user approvals allows the worker pool to process other jobs without blocking or leaving agents unmonitored.
• –Credential isolation: Injecting secrets via external proxies ensures unattended background agents never handle raw credentials directly.
• –Economics of long horizons: As frontier models become cheaper per solved task, stateful agents that maintain continuity across hours or days are finally viable in production.