What should an AI agent do when its AI goes down?

Key points

• When the AI account behind nox.cool refuses work, your task isn't marked as failed. It stays open, says why in plain words, and waits.
• nox.cool checks every couple of minutes, only while the AI is out. When a check works, every waiting task carries on by itself.
• A routine due in that time skips that one run, with the reason in its history. The skip doesn't count towards the failures that pause a routine.
• If the AI is only busy or out of reach, a task tries again 30 seconds, 2 minutes and 5 minutes later before giving up.
• Nothing about what Nox may do changes while it waits: OK cards still come first, and Stop still works.

nox.cool's AI agent, Nox, now keeps your task when its AI is unavailable. If the AI account behind nox.cool refuses work, your task isn't marked as failed. It says why in plain words, waits, and carries on by itself when the AI is back, with nothing for you to do. It's part of nox.cool's early access.

#Why would an AI agent stop working at all?

An AI agent is two things: the part that does the work (a browser, a computer, your files and apps) and the AI that decides what to do next. The second part is nearly always a service reached over the internet, and it can say no for reasons that have nothing to do with your request:

• It's busy. Too many people are asking at once, so it tells you to come back shortly.
• It can't be reached. The network fails, or the service itself is down.
• The account behind it is refused. A billing limit, a problem with a key, an account that has been switched off. This one doesn't clear in a few seconds.
• The request itself is wrong. Something in what the agent sent can't be read.

Those four look alike to a program, and they call for four different responses. An ordinary chat doesn't have to care much. If an answer fails you press the button again, and you've lost a few seconds.

An agent is different, because you've walked away. You asked it to work through a list of suppliers, and you went to a meeting. You have a routine that writes your morning briefing while you sleep. You find out what happened only later, and by then it's your problem to work out what went wrong.

#What usually goes wrong when the AI is refused?

Three things, and they make each other worse.

The message blames the wrong thing. The task says something like "internal error, start a new task". That reads as though your task broke, when nothing about it was wrong. You start it again, it fails the same way, and you have no way of telling whether to try once more or give up for the day.

Your work is thrown away. The task may have been twenty steps in, with a page open and half a list checked. Marked as failed, it's over. Starting again means starting from nothing and paying for the first twenty steps a second time.

Your schedule quietly falls apart. Sensibly, scheduled jobs usually pause themselves after a few failures in a row, so a broken one doesn't run for ever. But if the failures weren't the job's fault, a brief problem on the provider's side leaves all your routines switched off. They stay off after the service comes back, until someone notices the briefing has stopped arriving.

None of that is your doing, so none of it should be your job to clean up.

#What does Nox do now?

It tells the four cases apart, and treats each one the way it deserves.

The first row is the new one. Here is what it looks like, on a task to find three quiet hotels in Lisbon:

We chose those words with some care. "On our side" says whose problem it is. "Your task is safe" answers the first thing you'd worry about. "Nothing to do" saves you from wondering whether you should restart it. The usual "Thinking…" goes away too, because nothing is thinking, and an agent that pretends otherwise is worse than one that admits it.

#How does Nox know the AI is back?

The first refusal starts a small check. Every couple of minutes, nox.cool sends the AI the smallest request it can, one word, and only while the AI is out. If the answer is another refusal, it looks again later. If it works, three things happen together:

1. Every waiting task carries on from where it was. Steps it had already finished stay in its record, so it doesn't start over. What was held up was its next request to the AI.
2. Routines go back to normal at their next scheduled time.
3. The notice to nox.cool's team clears by itself.

You don't get a message asking you to come back, and you don't press anything. If you open the task later, the answer is simply there.

#What happens to routines?

A routine that comes due while the AI is out skips that one run. It isn't saved up to run later, and that's deliberate. A morning briefing delivered at four in the afternoon isn't a morning briefing, and three days of held-up checks arriving together are noise.

Three things matter about the skip:

• It's written down. The routine's history says "skipped: nox.cool's AI was unavailable", so a gap in your briefings has an explanation.
• It doesn't count against the routine. Routines pause themselves after repeated failures, and a skip isn't a failure. Nothing is switched off, and you have nothing to switch back on.
• The next run is normal. When its time comes, it goes ahead as usual.

#Who gets told?

We do, once. The nox.cool team gets a single notice when the AI account is refused, and it stays open until the account is working again. You don't get a stream of error messages about something you can't fix, and we don't have to hear about it from you first.

#What keeps you safe while a task waits?

Waiting doesn't loosen anything:

• Nothing happens in the meantime. A waiting task isn't browsing, clicking or sending anything.
• OK cards still come first. When the task carries on, anything important it wants to do still stops for your OK, with the exact details.
• Stop still works. If you've changed your mind while it waited, press Stop and it won't carry on.
• It costs you nothing. A waiting task isn't using the AI, and the small checks belong to us rather than to any task.

#What should you look for in any AI agent?

If you're choosing an agent to leave running while you get on with your day, it's worth asking how it behaves on a bad day, not only a good one. These are the questions we'd ask of any of them, ourselves included:

#How did we test it?

We ran nox.cool against a stand-in for the AI that refuses requests the way a real account does, then started answering again: 17 checks, covering tasks that wait, routines that skip, the single notice, Stop, and everything carrying on by itself. We also confirmed that a real refusal from the AI service is recognised for what it is. So far, the moment of coming back has been tested with the stand-in, and it hasn't yet been seen with a real account.

#Questions and answers

What should an AI agent do when its AI service is down?

Keep the task, say plainly what's happening, stop calling the service, and carry on by itself when the service is back. nox.cool's AI agent, Nox, does this when the AI account behind it refuses work, so a problem on our side doesn't become a failed task on yours.

Do I have to do anything when Nox is waiting for the AI?

No. The message says so: your task is safe and carries on by itself. You can press Stop if you no longer want it done.

Does Nox start over when it carries on?

It doesn't start over. The task carries on from where it was, with the steps it finished already in its record. What was held up is the next request to the AI.

Does a waiting task cost me anything?

No. A waiting task isn't using the AI. The small checks that notice when the AI is back belong to nox.cool and aren't charged to any task.

What happens to my scheduled routines during an outage?

A routine that comes due while the AI is out skips that one run. It isn't saved up for later, its history gives the reason, and it doesn't count towards the failures that make a routine pause itself. The next scheduled run goes ahead as usual.

Does every kind of failure wait like this?

No. An account refusal waits for as long as it takes. A busy or unreachable service gets three more tries, minutes apart. Anything else is reported as a failure, with what to do next, because waiting wouldn't fix it.

Where can I try it?

It's part of nox.cool's early access, which is by invite for now.