On 13 November 2026, Anthropic published a guide on agents that keep working when nobody is in the chat. Always on is the experience, not staying awake all night. The agent sleeps, wakes for a short run, saves what happened, and goes back to sleep. Hours it saves are capacity, not cash. Unattended agents rarely go rogue. They stall, stop, or break silently. Run an agent attended first, then let it run on its own, and have a second agent check the work. If a team or vendor cannot say how the inbox, timers, memory, permissions, and outbox work, the agent is not ready. The later test: would you know if it stalled overnight.
Key takeaways. Always on is the experience, not the runtime. Hours saved are capacity, not cash. Run attended first, then unattended, with a second agent reporting success and failure. Automatic replies are a comfort-level choice. Automatic drafting is a different decision. Start the agent quiet, between over-notification and missing what you need to know.
Five building blocks (Anthropic).
- Event inbox. Every trigger writes to one durable queue, with one active worker. If the worker dies, a new one can retry. Ask: does every trigger go through one queue, and what happens if the worker dies mid-run.
- Timers. Follow-ups the agent sets for itself, and recurring schedules. Ask: can I see them, and can I turn them off.
- Memory. Keep only what it cannot look up again. Save a pointer if the fact already lives elsewhere. Ask: what does it keep, where does that live, and can we read it.
- Permissions. A run holds a person's access. Each action should proceed, block, or wait for a named person's approval. A proxy outside the agent adds the password or the key. Ask: does the agent hold the credentials, which actions wait, and what if that person does not answer.
- Outbox. The user gets the outcome and a few details, not a live stream of every step. Ask: how does it decide when to message someone, can we start it quiet, and is there a log of sent and unsent messages.
Checklist (as spoken). You do not need to write this down as you listen.
- Run the agent attended first. Then let it run on its own. Why: you learn how it behaves while someone is still watching.
- Have a second agent check the work and report success and failure. Why: they stall, stop, or break silently, and the job can look like it is still running.
- Write down which messages can go out on their own, and which wait for a person. Automatic replies are a comfort level. Automatic drafting is a different decision. Why: drafting can go first. Sending is still a separate yes.
- Ask the five questions above. Why: if a vendor cannot answer them, the agent is not ready to run unattended.
- Once it is running, ask whether you would know if it stalled overnight. Why: that is a better test than asking whether it ever went rogue.
Terms. Always-on agent: feels unattended, but sleeps between short runs. Memory: what is kept between runs, and is inspectable. Credential proxy: inserts the password or key. Outbox: how the outcome reaches a person. Start quiet: fewer messages first, then earn attention over time.
Examples named. We run agents unattended after an attended check, with other agents reporting success and failure. They rarely go rogue; the watch is for silent stall. Some messages get automatic replies, many do not. Most emails get an automatic first draft. Drafting is not sending.
Sources: Anthropic guide; Elvis Saravia on X.
The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.