Troubleshooting
What actually stops when you close your laptop
Your agent was supposed to send the report at 9am. It is 11am, nothing arrived, and the logs stop at the moment you shut the lid last night. This page is the diagnosis, not a pitch: the seven things that need a live process, how to work out which one bit you, and what you can genuinely fix yourself.
The short answer
An agent is a process. When a machine suspends, the process stops. It does not crash and nothing is corrupted - it simply is not running, and anything that depended on it being awake did not happen.
The reason this is confusing is that the failure is invisible. There is no error, because nothing errored. There is just an absence where a thing should have been.
Seven things that need something awake
1. Scheduled runs. A schedule fires on a clock. If nothing is listening at 9am, 9am passes. Crucially, missed runs are not queued and replayed when the machine wakes - the scheduler is not working through a backlog, it is watching a clock.
2. Standing instructions. Anything phrased as "keep an eye out for" is implemented as a poll on an interval. No process, no poll, and no record afterwards that the check was skipped.
3. Inbound messages. A customer writing on WhatsApp or Telegram is not lost - the platform holds it - but nobody replies until something is listening again. On a customer channel, a nine-hour reply time is the part they remember.
4. Long-running work. Reading a folder of documents, or a research task with many steps, takes longer than a lunch break. Suspending the machine interrupts it rather than pausing it.
5. Inbound endpoints. Webhooks and shared links point at an address. A laptop that changes network - office to home, wifi to tethering - changes or loses that address, and the far end gets a connection error it will not retry forever.
6. Connector renewals. The one nobody predicts. Access grants to connected accounts expire on their own schedule and have to be renewed ahead of expiry. A machine that was asleep through the renewal window wakes up to a connection that now needs a human to re-authorise it.
7. Anything another person was relying on. Not a technical category, but the one with consequences. If a colleague or a customer expected an answer, the outage was theirs, not yours.
Working out which one bit you
Match the symptom rather than reading logs first.
Nothing at all happened overnight and everything works now you are back: the process was simply down. Nothing else to find.
Everything works except one integration, which now says it is not authorised: that is the renewal trap in point 6. Reconnect the account; it will not fix itself, because consent cannot be renewed automatically once it has genuinely lapsed.
The agent replies to you but a customer says they got nothing: the channel connection is up but was down when they wrote, and their message arrived to nobody.
A task shows as started and never finished: it was interrupted mid-run. Start it again rather than waiting.
What sleep does not damage
Worth stating plainly, because the common fear is the wrong one: your data is fine. Conversation history, memory, files and configuration all survive a suspend, a shutdown and a restart. Nothing needs rebuilding. The thing you lost was time, not state.
The three local workarounds, honestly
Stop the machine sleeping. Change the power settings, or keep a lid-open utility running. This works, and it turns a laptop into a space heater that must never be carried anywhere. It also does nothing about the network changing.
Use a spare machine. An old desktop in a cupboard is a genuinely good answer, and plenty of people run this way for years. You are now responsible for its power, its network, its updates and its disk.
Rent a small server. A few dollars a month buys a machine that does not sleep, a stable address, and a process that comes back after a reboot. This is what most self-hosting communities recommend and it is correct advice.
All three fix availability. None of them changes who is on call when a platform changes its rules on a Sunday.
When to stop patching
The point to stop is not a feature threshold, it is a moment: when someone other than you is depending on the agent being there. A personal assistant that is down for a day is an inconvenience you noticed. An agent answering customers that is down for a day is a business problem you did not notice, which is worse.
FAQ
Why didn't my AI agent run last night?
Almost certainly because the machine it runs on was asleep. A schedule fires on a clock and needs a live process at that moment; if nothing is running, the time passes and nothing is recorded. The absence of an error is the clue - nothing failed, because nothing ran.
Will a missed scheduled task catch up when my computer wakes?
No. Missed runs are not queued and replayed. The scheduler watches a clock rather than working through a backlog, so a 9am run missed while the machine was asleep does not fire at 11am when you open the lid. Start it manually if it still matters.
My integration says it is not authorised any more. Is that related?
Very likely. Access grants to connected accounts are renewed ahead of their expiry, and that renewal needs something running. A machine asleep through the renewal window can wake to a connection that now needs a human to reconnect it. It will not recover on its own, because consent cannot be re-granted automatically.
Did I lose any data?
No. Conversation history, memory, files and configuration all survive a suspend, a shutdown or a restart. Sleep costs you availability, not integrity.
What is the cheapest way to keep an agent running 24/7?
A small rented server, which most self-hosting guides put at a few dollars a month. That fixes uptime completely. What it does not change is that you are now the operator - updates, expiring tokens and platform changes all land on you, at unscheduled times.
Related reading
Where to go next.
Local vs cloud AI agents
The decision version of this page, once you know which thing broke.
Always-on AI agents: what 24/7 requires
The five mechanisms unattended work depends on, and their limits.
Olano vs OpenClaw
If you are running a self-hosted harness and weighing the move.
Docs: automation
How scheduled work and standing instructions are configured.
If it mattered enough to notice, it matters enough to fix properly.
Tell us which run was missed and who was waiting for it. We will say whether a ten-dollar server solves it or whether the obligation has outgrown a machine you carry around.