On November 13, Anthropic published a post on the official Claude blog breaking down the architecture of always-on agents. An always-on agent does not actually run around the clock: it mostly sleeps, wakes up when an event arrives, does its work, and goes back to sleep. Better model judgment and falling inference costs are moving this kind of agent from demo to practical use — on long-horizon tasks, Opus 5.5 costs 78% less per solved task than Opus 5. Anthropic says customers are already asking how to build this experience into their own products.
What each of the five primitives does
Always-on agents share the same core loop as ordinary agents; the difference is the surrounding architecture, built from five primitives. The event inbox handles receiving: webhooks, scheduled tasks, and background jobs all write events to a database table first, where they are deduplicated before a worker picks them up — so two events arriving at once cannot start competing runs, and a crash cannot lose an event. Timers handle proactivity and come in two kinds: follow-ups, deadlines the agent sets for itself, and schedules, recurring work on a fixed cadence such as a daily newsletter, fired by a scheduler that polls on an interval.
Memory keeps continuity across many separate runs: it stores only what the agent cannot look up elsewhere, in four categories — user instructions, user notes, messages already sent, and task progress — and Anthropic's own products implement it as a directory of files. Permissions are the riskiest part: an unattended run holds the user's credentials, so Anthropic recommends injecting credentials through a proxy before requests go out, blocking high-risk actions in auto mode, and deferring tool calls until the user confirms. The outbox paces results to the user: too noisy and it disturbs, too quiet and it misses things. Anthropic recommends starting quiet and gradually earning more of the user's attention.
Who can use this, and where the limits are
This architecture suits teams that want to make an existing agent proactive or build an always-on experience from scratch; typical cases are customer follow-ups, scheduled briefings, and unattended background tasks. But it answers when the agent wakes, what it remembers when it does, and how it reports back — not the model's capability ceiling. Sessions that grow too long still need compacting or restarting, and anything that must persist has to be written to memory in time. Every run also carries context overhead, so Anthropic recommends keeping the initial context lean and using prompt caching to hold costs down.
Anthropic positions the post as a build guide, not a product launch: no new product or launch date was announced. Developers can use the five primitives to audit their own architecture for gaps, and how agent protocols work in practice is worth reading alongside it.
