Anthropic's Blueprint for Always-On Agents: Five Building Blocks, and Start Quiet

An always-on agent is not literally always on. In Anthropic’s engineering guide to building always-on agents, these stateful, proactive agents sleep and wake for discrete runs, carry context across sessions and act without direct supervision. Anthropic credits the renewed interest to better model judgment about user intent and to cheaper intelligence: on one long-horizon coding benchmark, Claude Opus 5.5 cost 78% less per solved task than Opus 5, against 40% less on typical workloads.

The guide reduces the architecture to five building blocks. An event inbox — usually a database table with a handler and a single worker — takes webhooks, schedules and background jobs, removes duplicates, batches pending events and survives crashes. Timers, stored separately and fired by a scheduler, cover follow-ups the agent sets for itself and recurring schedules. Memory lives in durable storage and holds only what the agent cannot look up again: user instructions, notes about the user, past messages and per-task progress files. Permissions combine a credential proxy outside the agent’s environment, classifier-based auto mode for higher-risk actions, and deferred tool calls that pause a run until the user approves. An outbox and dispatcher pace what reaches the user.

That last block carries the guide’s sharpest advice: agents should “start quiet” and earn attention, with the volume tuned through evals.

The permissions block matches what Anthropic is shipping for regulated buyers. Its self-hosted Claude Managed Agents private beta keeps credentials out of the sandbox, can limit network access to an allowlist of hosts, and records every tool call and result for export over OpenTelemetry. Anthropic says the demand comes from agents doing unattended work such as fixing broken builds, testing releases and triaging alerts.

For teams scoping a first proactive agent, the five blocks double as a build checklist. Without an inbox, events get dropped or duplicated; without a dispatcher, the agent talks too much. Decide what it may do unattended, what pauses for approval and how quietly it starts — before deciding what it does.