Attention needs admission control

The useful starting point is narrower than the slogan “attention spans are shrinking.” The logged unit is time on one screen before switching. It is not a stopwatch on intelligence, motivation, or the maximum duration of concentration. That distinction matters because a switch from an editor to a terminal may preserve one project, while a switch from an architecture problem to an unrelated message can evict the state that made the work possible.

The measurements still expose pressure. Mark reports that average time on one screen fell from roughly two and a half minutes in 2004 to 75 seconds around 2012 and about 47 seconds in more recent logging, with a 40-second median. Her UC Irvine interview separately describes about 25.5 minutes before people return to an original interrupted project after cycling through other work. The second number is not “time to regain focus”; it is elapsed time until the original project is picked up again.

The engineering failure is uncontrolled admission. Email, chat, meetings, alerts, browser tabs, and internal urges all behave like producers. The current task is a single consumer carrying expensive in-memory state: hypotheses, call paths, rejected approaches, constraints, and the next uncommitted step. Let every producer interrupt immediately and priority collapses into arrival order.

A defensible policy separates arrival from admission. Ordinary messages may arrive without being scheduled. Incident channels can preempt because their cost of delay is explicit. Everything else waits for a breakpoint or a planned drain window. This is not an attempt to maintain nonstop focus; Mark’s interviews emphasize attentional rhythms, replenishing breaks, and individual differences. The goal is to choose switches instead of accepting every switch at producer speed.

The queue analogy becomes misleading if it is mistaken for brain anatomy. Petersen and Posner describe attention as interacting systems: alerting prepares the organism for a signal, orienting prioritizes a source or location, and executive networks regulate conflict and ongoing control. Later work in the review separates dorsal and ventral orienting systems and distinguishes cingulo-opercular from frontoparietal control. There is no tiny scheduler in the head assigning CPU slices.

That distributed picture explains why a salient arrival and a deliberate switch are not the same event. An external cue can reorient processing before the current objective has been safely checkpointed. Laboratory task-switching research adds a narrower result: responses are typically slower and more error-prone immediately after a task change, and preparation reduces but does not usually erase the switch cost. Reviews attribute the cost to a mixture of task-set reconfiguration, interference, and carry-over rather than one universal recovery timer.

The practical consequence is modest but useful. Reduce unnecessary orienting cues, prepare intended switches, and externalize the current task set before changing objectives. Do not claim that a particular app notification “hijacks” one named brain region or that a fixed number of minutes applies to every switch; the evidence supports distributed control and measurable switch costs, not a cartoon circuit diagram.

Attention is not philosophically interesting only because processing capacity is limited. The Stanford Encyclopedia of Philosophy frames it as the selective directedness of mental life, then maps disagreements about what makes that selection possible: capacity limits, the need to sustain one coherent course of action, precision-weighting, or competition between streams. Attention can be perceptual, intellectual, or action-oriented; it can be voluntary or captured by salience.

William James made the stronger experiential claim in 1890: what enters experience is not everything presented to the senses but what selective interest gives foreground and background. On that view, an interruption policy does more than improve throughput. It helps govern which questions, people, obligations, and possibilities repeatedly become vivid enough to shape the working self.

This introduces an autonomy test for interface design. A system that makes every arrival equally salient does not merely consume time; it delegates selection to senders, ranking functions, and notification defaults. A respectful tool lets the person notice why a demand is salient, delay it without losing it, and return to a chosen object of thought. Collective quiet windows then become more than a productivity hack: they protect a shared ability to decide what deserves concern.

My experience is what I agree to attend to. Only those items which I notice shape my mind—without selective interest, experience is an utter chaos.

William James, The Principles of Psychology, Chapter XI (1890)

A breakpoint is where the current task can be reconstructed cheaply: after a test turns green, after a paragraph closes, after a trace is captured, or after the next action is written down. Before opening a communication queue, checkpoint the active state in one or two lines. The note should name the objective, the last verified fact, the next action, and any command or file needed to resume. This externalizes the state most likely to be lost during the switch.

Then drain deliberately. Time-box ordinary channels, answer what can be closed, turn larger requests into scheduled work, and leave the queue again. A sender should not be able to convert an asynchronous medium into a synchronous obligation merely by sending another ping. Conversely, a real incident needs a separate channel with a documented severity threshold so silence does not hide urgent operational risk.

Individual notification settings cannot repair a team that rewards immediate replies. Mark argues for collective quiet windows because one person who disconnects can otherwise be penalized by the surrounding communication norm. Microsoft’s telemetry shows why this matters: for the highest-volume fifth of users, core-hours pings land every two minutes, while half of meetings occupy two periods that often overlap reported productivity peaks.

The team-level contract should publish response-time expectations by channel, reserve one path for real incidents, protect quiet windows, avoid booking over every common focus period, and judge work by outcomes rather than visible reactivity. Agents can help summarize queues and prepare checkpoints, but using AI to answer every message faster only accelerates the same broken scheduler. The leverage comes from reducing unnecessary admissions, not increasing interrupt throughput.

1. Name the urgent channel and define what qualifies; everything else is asynchronous by default.
2. Publish realistic response windows so silence is not mistaken for neglect.
3. Create shared quiet blocks and keep routine meetings outside them.
4. Checkpoint long-running work before switching; make resumption state part of the workflow.
5. Review interruption sources as a team metric, not an individual willpower defect.