Always-On AI Agents Wake for Events, Then Return to Sleep

    Always-On AI Agents Wake for Events, Then Return to Sleep

    Always-on AI agents are gaining renewed attention as newer models become more capable of working independently and less expensive to run. The term describes an experience, not a system that computes without pause: these agents sleep between jobs, then wake when a message, scheduled task, webhook, or other event gives them something to do.

    The underlying agent loop is familiar, but the infrastructure around it’s different. It must decide what triggers a run, preserve useful context between runs, control what actions the agent can take, and deliver results to the user. Anthropic describes that loop in its guide to building effective agents.

    Model improvements have helped make the approach more practical, according to Anthropic. Newer models are better at interpreting intent and deciding when to contact users, while lower costs make extended tasks more affordable. On one long-horizon coding benchmark, Opus 5.5 cost 78 percent less per completed task than Opus 5; across typical workloads, the reduction was 40 percent.

    Each period of activity is a run. An event starts the run, the agent performs its work, and information about the outcome is saved. The run takes place within a session, which contains the conversation and tool calls. When the work is complete, the agent returns to an idle state until the next event arrives.

    That cycle relies on five building blocks: an event inbox, timers, memory, permissions, and an outbox. The inbox collects triggers rather than sending each one directly to the agent. A durable record of incoming events lets a background worker retrieve and process them, even if the worker fails during a job. It also helps prevent overlapping runs from duplicating work or overwriting one another.

    In a typical setup, an event handler quickly records incoming requests, while a worker retrieves pending items and starts a run. Events that continue an existing conversation, such as a user’s reply, can resume its session; unrelated events can begin new sessions. The worker can also combine a burst of messages into one run. Keeping a single active worker for each inbox helps avoid conflicts, while event identifiers can be used to filter out duplicates.

    Timers give agents a way to act later. They can represent one-time follow-ups, such as checking whether someone has replied, or recurring jobs, such as preparing a daily update. A scheduler checks a durable timer store, marks completed follow-ups or advances recurring schedules, then places due work in the event inbox. The timer itself does not run the agent.

    Memory provides continuity across separate runs. It should live in durable storage so it remains available after a worker shuts down. Useful material can include a user’s instructions and preferences, a record of what the agent has already said, and progress notes for ongoing tasks. Information already available in another system, such as an issue tracker, usually needs only a reference rather than a copy.

    At the start of a run, the agent loads its system instructions, tools, and relevant memory. Keeping this baseline context small can reduce repeated costs, and prompt caching can help. An existing session may also restore earlier messages and tool calls. If a session grows too long, older content can be condensed or a fresh session started, but details that must survive either change need to be saved separately.

    Memory can also be reviewed across past sessions to update stale notes or identify patterns that emerge over time. This process, sometimes called “dreaming,” is distinct from saving information as a run happens: it periodically curates what the agent should retain.

    Permissions are a central safety concern because an unattended run may act using a user’s access. Systems can keep credentials outside the agent and add them to requests through a proxy. They can also apply risk checks that block dangerous actions, while reserving important decisions for human review.

    When an agent needs clarification or approval, it can defer a tool call rather than remain stuck in a live run. The system saves the session, presents the request to the user, and treats the reply as a new inbox event. A later run can resume the work, while the worker remains available to handle other events in the meantime.

    Most conventional agents stream their work as it happens. Always-on agents tend to operate in the background and send a result, a request for input, or nothing at all if there is no useful update. A formal outbox can store outgoing messages for a dispatcher to deliver, helping systems manage delivery and recover from errors. Teams also need to tune how often the agent contacts people: too many messages become noise, while too few can leave users uninformed.

    Together, the five components create a repeating loop. An event wakes the worker, which starts a run in a new or existing session. The agent consults memory, acts within its permissions, saves new information, and may queue a message for delivery. It can also set a timer for future work. Each run is finite, but the cycle continues as new events arrive, creating the impression of an agent this is always available.

    Leave a Reply