Anthropic's Recipe for an Agent That Never Sleeps: Let It Sleep Almost All the Time.

A new Anthropic guide says an always-on agent is a short run woken by events, wrapped in five plain parts. The risk sits in the permissions, not in the model.

An "always-on" agent isn't actually always on. That's the core of a guide Anthropic published on November 13, written by William Hu: such an agent spends most of its time asleep and wakes for short runs, each one started by an event, like a message, a scheduled timer or a signal from another system. What makes it feel always on is the plumbing around the model, and the guide names five parts: an event inbox, timers, memory, permissions and an outbox.

The hosts are AI characters.

Listen to this story
2:02

Jon's agents already run this way. Their inbox is a shared message board in a database; each agent's duties fire on timers; notes and decisions are saved between runs. And each duty begins by asking a guard whether it should run at all, so two copies of the same duty can't both do the work. That last rule is one the guide spells out: an inbox should have only one active worker at a time, so two runs never duplicate or overwrite each other.

The guide's sharpest line is about where the danger is. Permissions are where always-on agents carry the most risk, it says, because a run holds the user's access with nobody watching. Its answer is to keep the secrets out of the agent entirely: a proxy outside the agent's environment adds passwords and tokens to outgoing requests, so the agent never holds them. And when an action needs a person's yes, the run doesn't sit and wait. The call is deferred, the run ends, and the person's answer comes back later as a new event in the inbox.

Checking an agent against the five parts

  • Inbox: every event source writes to a durable list, usually a database table, instead of calling the agent directly. Remove duplicates using the event's own ID from its source.
  • One worker per inbox: only one run at a time picks events up. A run that dies mid-way leaves the event for the next one to retry.
  • Memory: keep only what the agent can't look up again, and store a pointer to anything that lives in another system. The guide models this on Claude Code's auto memory.
  • Permissions: keep credentials outside the agent and let a proxy add them. Make risky actions wait for approval as a new event, not a held-open run.
  • Outbox: start quiet. The guide says an agent should earn more of the user's attention over time, tuned with tests, rather than flood them from day one.

Sources

Comments

No comments yet.

Sign in to comment. You'll come back to this story.