A notification product is in the business of interrupting you. That is worth saying plainly, because the strongest argument against building one is also the most honest. Interruptions are expensive, and the research that measures them does not flatter the person sending them. So before the design, the cost. What an interruption actually costs In a field study of information workers, people spent only about eleven minutes on a piece of work before they switched or were pulled off it. When they returned to an interrupted task the same day, getting back to it took an average of twenty-five minutes and twenty-six seconds. Not twenty-five seconds. Twenty-five minutes, spent on the return alone. The cost is not only time. People interrupted while they work tend to finish faster, speeding up to compensate, with no drop in quality. They pay for it elsewhere, in more stress, more frustration, more effort, more time pressure. And it lands quickly. After only about twenty minutes of working under interruption, people reported significantly higher stress and mental workload than people left alone. Put those together and you have the real shape of an interruption. It is cheap to send and expensive to absorb. The sender sees one notification. The receiver pays in minutes, and in the quieter currency of effort and irritation. An interruption is cheap to send and expensive to absorb. The whole design problem is making sure the few you send are worth what they cost. Which is why we send so few This is why Seorak does not stream activity at you. It watches a session and stays quiet, then surfaces a small set of watch notices only when one crosses into territory worth the cost. A cost spike. A retry loop circling the same fix. A run going far longer than the work seems to warrant. The state crossing into needs-you. Most sessions trip none of these, and you hear nothing. That silence is the feature. Roughly, the line we are trying to draw. Treat it as illustrative, not a precise rule. Seorak stays quiet, Seorak interrupts a long run still moving toward a commit, a run stuck in a retry loop, the clock outrunning any commit token count climbing on changes that are landing, a cost spike with nothing landing under it a session winding down on its own, the state crossing into needs-you When a watch limit trips, it says what it saw, plainly, and gets out of the way. "Your agent has been retrying the same fix for 25 minutes" is a sentence you can act on in a second. That number is an example, not a threshold we are quoting. The aim is that the interruption pays for itself: you learn something you could not have learned by waiting, at a moment when you can still do something about it. We would rather miss one than spend your attention on noise. A product that nudges you about everything teaches you to ignore it, and a notification you have learned to ignore costs you the interruption for nothing. The part about timing we have not earned yet There is a better version of this, and the research describes it. Interrupting at a natural breakpoint, the seam between two chunks of work, instead of mid-thought, is markedly less costly. Less annoyance, less frustration, less mental effort, for the very same message. And those breakpoints can be spotted automatically, from patterns in activity alone, without the system needing to know what task you are on. Here is the honest limit. Seorak does not do that. It fires on the recorded state, not on your cognitive state. It knows the run went to stuck and the clock has outrun any commit. It does not know whether you are mid-thought or between tasks, whether this is a good second to look up or the worst one. The timing it can offer is the timing of the session, not the timing of your attention. So the discipline we actually have is rarity and a high bar. Interrupt seldom, and only on a reading strong enough to be worth the twenty-five minutes it might cost you to come back. Detecting the right moment inside your own work is a real, studied capability. It is also one we have not built, and we would rather say so than imply the nudge knows more about your focus than it does. The summary still waits at the end of the session, and it costs nothing to read later. The interruption is the part we have to ration. Every one we send is a small bet that what it shows you is worth more than the time it takes you to get back.