Anthropic has published a guide to building always-on agents, and its central claim is that the name is misleading. Mostly, the agent is idle. What the user experiences as a constant presence is a series of short runs, each triggered by an event, finished, and logged before the agent goes dormant again.
The guide, written by William Hu and posted on Anthropic’s claude.com resources pages on November 13, matters because customers keep asking the company how to build this. Anthropic says some want existing agents to take more initiative, and others plan to start fresh. The post is its answer to both groups, and it reads as a reference architecture rather than a product announcement.
Anthropic ties the renewed interest to two shifts in the models. The latest ones are better at judging user intent and at knowing when to check in. And cost is falling, especially for long tasks. Anthropic cites a long-horizon coding benchmark where Opus 5.5 was 78% cheaper than Opus 5 for each task solved, versus 40% cheaper on typical workloads. Those figures are the company’s own, and the post does not name the benchmark.
The architecture rests on five parts. The first is an event inbox: a durable table where webhooks, scheduled jobs, and user messages land instead of calling the agent directly. Direct calls break in two ways, Anthropic notes. Simultaneous events can launch rival runs, and a failure mid-run can drop an event entirely. With an inbox, a single worker per inbox picks up pending items, batches them (several rapid user messages become one reply), and marks them done afterward.
Timers form the second part. Anthropic splits them into follow-ups, which are deadlines the agent sets for itself, and schedules, which recur on a known cadence. They sit in their own store, apart from the inbox, and a scheduler polls that store and drops an event into the inbox when something comes due. The guide says teams can build that scheduler or lean on a durable execution engine.
Memory is third, and the guide is blunt about scope: store only what the agent cannot look up again. If a fact lives in an issue tracker, keep a pointer. The recommended files cover standing user instructions, notes on the user, what the agent has already said (so it does not repeat itself), and a progress file for each ongoing task. Anthropic also floats “dreaming,” a pass that reviews old sessions, replaces stale entries, and spots patterns across many runs.
Permissions carry the most risk, in Anthropic’s telling, because an unattended run operates with the user’s own access rights. Its recommendations: keep credentials in a proxy that sits beyond the agent’s sandbox and injects them into outgoing requests; use classifier-based review of risky actions, as in Claude Code’s auto mode; and defer high-stakes calls. A deferred call ends the run, saves the session mid-call, shows the user a request, and treats the answer as a new inbox event. Meanwhile the worker keeps handling other events.
The outbox is the fifth part. A run can end by sending a message, by deciding nothing is worth sending, or by deferring a call for approval. Most teams can get away with a simple send tool, the guide says, and need a formal outbox only to pace messages or recover from delivery errors. Anthropic advises building evals for message volume and starting quiet, so the agent earns attention over time.
The unspoken market angle is that none of this is exotic. An inbox table, a timer table, a memory directory, a proxy, and an outbox table are ordinary backend parts, which means the moat for always-on products will sit in judgment and tuning, not in plumbing. That also explains why Anthropic is publishing the recipe: more agents running on its models means more runs billed.
For any team planning a proactive agent this quarter, the first decision is the notification budget, since a noisy agent gets muted before it ever proves useful.
Based on “Building always-on agents” by William Hu, published by Anthropic on claude.com on November 13, 2026.