An agent that needs you constantly is annoying. An agent that needs you rarely is the one to watch. The better it runs on its own, the less you check it, and the further you drift from the state of the work. Then the single moment it actually needs you arrives, and you are not there. You have to find your way back into a session you stopped following an hour ago. Automation's oldest irony This is not a failure mode agents invented. It is the oldest problem in automation, and it has a name. Decades of research on automated cockpits, plants, and control rooms keeps returning to the same irony: the more capable the system, the more crucial the human's remaining contribution may become, not the less. You automate to take out the unreliable human, and you are left with a human whose one remaining job, catching the rare moment it goes wrong, is the job people are worst at. Monitoring for something that almost never happens is close to impossible to do by attention alone. The same literature is blunt about it: a person cannot hold effective attention on a source where very little happens for much more than about half an hour, so the watching has to be handed to an automatic alarm rather than to willpower. That is the part worth sitting with. The answer the research reaches for is not try harder. It is build the alarm. The cost is not only the missed moment. It is how slow you are once you do notice. When people monitor an automated process passively instead of steering it, their awareness of the situation decays, and that decay shows up as slower decisions at exactly the point a failure forces them to take back over. The same work found that keeping the operator in partial control, instead of automating them out entirely, preserved more of that awareness. Drifting all the way out of the loop is what makes stepping back in hard. It is tempting to read this as a discipline problem, the kind experience fixes. It is not. The drift is an attention-allocation failure that shows up under load, when other tasks compete for the same attention the automated one needs, and it has been found in expert participants as well as naive ones. It does not go away with practice. So keep a closer eye on it is not a plan. It is the very thing the research says people reliably cannot do. Where the nudge comes in This is the reason attention tooling exists at all, and it is a narrower claim than we send notifications. Seorak watches a session for a small set of conditions and speaks up when one crosses into territory worth your attention: the state moving to stuck or needs-you, a retry loop circling the same fix, a cost spike, a run going far longer than the work warrants. It does this without reading your code or your prompts. The stats are derived on your machine, and only those leave. When one fires, you get a plain line on whichever surface you are near, terminal, web, or phone, while there is still a session to change. A better agent does not need less watching. It needs watching you can no longer do by hand. The autonomy is what creates the need for the nudge. The alarm has to earn it There is a tension here we take seriously. The thing that corrects an out-of-the-loop problem is an alarm, and the same research is just as clear about how alarms fail. One that goes off when nothing is wrong gets ignored, and once it is ignored it protects nothing. The studied failure of automation is not only overreliance on it. It is also the alarm that cried wolf until people tuned it out. So Seorak is built to under-alert. A small set of alerts, sent rarely, should stay quiet rather than teach you to swipe them away. The economics that justify the nudge are the same economics that demand it be rare. What this cannot do Here is the line it does not cross. Seorak's nudge is an alert, not situation awareness. It can tell you a run looks stuck, or has gone far longer than the work warrants, but it cannot hand back the understanding you lost while you were not watching. Re-reading the diff, deciding whether the agent is wrong, choosing whether to step in, that is still yours, and it still takes the minutes it takes. The alarm shortens how long you stay out of the loop. It does not restore the loop for you. And it is a heuristic. The readings are a strong tell, not a verdict, and none of the research here proves that Seorak fires at the right moment. We would not claim it does. What that research establishes is narrower and sturdier: the out-of-the-loop problem is real, it gets worse as the agent gets better, and the sane answer is an alarm built to respect the moment it interrupts. That is the problem Seorak is pointed at. Whether the nudge lands well is something your own sessions, not a citation, get to decide.