Context Switch

In a study published in 1972, sixteen preschoolers waited with the snacks hidden away under a tin, free to end the wait at any moment and take the one they liked less. The children told to think about the reward gave up in under a minute. The ones told to think about “fun things” waited about fourteen and a half minutes, and the ones given no instruction waited nearly thirteen. That matched Walter Mischel’s earlier runs, where children with the reward in plain view and nothing to distract them lasted less than a minute. The marshmallow test began as a study of the arrangements that make waiting easier.

The prediction it became famous for rested on thirty-five children in one condition, whose waiting correlated with their later Scholastic Aptitude Test scores. A follow-up of 552 children whose mothers hadn’t completed college found the link with achievement at fifteen shrank once family background, early cognitive ability and the home environment were accounted for. Adding the children’s other measures at four and a half left almost nothing.

Waiting also tracked the adult running the test. Children waited about twelve minutes for an experimenter who’d kept earlier promises and about three for one who hadn’t.

Self-control rated over years does predict a life. The Dunedin study followed 1,037 children and combined reports from parents, teachers, observers and the children themselves across years. Lower childhood self-control predicted worse health, poorer finances and more criminal convictions at thirty-two, after accounting for childhood intelligence and social class. Those ratings covered impulsivity, persistence and attention in daily behaviour, which is a broader thing than any laboratory executive score.

Executive function is the ninth function on this panel, and the panel gives it the chipsetchipsetThe support chips around a computer’s processor, connecting it to memory and devices. Their job is to move data between parts of the machine.. That choice is half-right. Control over everyday behaviour is real enough to predict health and money decades later, yet the laboratory version of it keeps dissolving into other abilities whenever researchers try to isolate it, and no chipset ever decided which task ran.

Executive function is the umbrella term for keeping behaviour pointed at a goal when an easier response competes with it. Miyake et al. gave the field its three-part account in 2000 by checking which executive tasks ranked the same students alike. Updating refreshes what’s held in mind and overlaps heavily with working memory. Shifting changes the rule in use. Inhibition stops a dominant response. The three were related but separable, which the authors called “unity and diversity”, and they never presented three as the full list.

Inhibition was the first to lose its own identity. Friedman et al., working with twins, separated what all the executive tasks shared from what belonged to each component. Updating and shifting kept something of their own. The inhibition tasks had nothing left once the shared ability was taken out. People still stop actions; what vanished was a separate ranking of who does it best.

The shared ability looked 99% heritable in that sample of seventeen-year-old twins. HeritabilityHeritabilityThe share of differences between people that a study attributes to genetic differences in that population. It describes variation across a group, and says nothing fixed about how far one person can change. came out at 46% in middle-aged male twins and 72% in a combined twin and adoption study of adults, with different tasks and populations each time. Updating tracks intelligence closely, and the shared ability only partly.

The three-part structure itself moves between studies. A review of forty-six samples, with a reanalysis of the twenty-one that could be refitted, found no model that fitted all of them. Children more often looked like one broad ability; adults more often showed a shared core plus distinct components. Another study, with reliable measures throughout, found no coherent executive-control ability at all.

No chip in a computer decides which task runs next. The support chips move data between the processor, memory and devices without knowing whose data it is, and most of their work has since moved into the processor itself. The choice belongs to the operating system’s schedulerschedulerSoftware that chooses which waiting task gets the processor next. It applies rules about priorities and when running work should give way., software that runs on the same processor as the programs it schedules. Each context switchcontext switchThe handover from one running task to another. The computer saves the current task’s state so it can resume and loads the next task’s state. costs time, most of it spent reloading data the previous task pushed out of fast memory.

People pay switching costs too, of two kinds. Changing rules from one trial to the next costs time, and so does working through a block where either of two rules might be needed, the global switch costglobal switch costThe extra difficulty of working in a setting where either of two task rules may be needed, compared with using one rule throughout. It includes keeping both sets of instructions available.. Only the second grows with age beyond general slowing. Nobody has shown that people pay theirs the way a computer does, by reloading what was pushed out.

Inhibition had already dissolved into shared control. In one study of nine executive tasks, shared control dissolved in turn into how quickly people took in information, with nothing specifically executive left over. That study wasn’t preregistered. Another found that executive tasks owe their link with intelligence to processes present even in the easy conditions, before the executive demand is added.

Frontal-lobe damage lowers scores on executive tests, which made them look like frontal-lobe tests. A meta-analysis of lesion studies on card sorting, word fluency and the Stroop test found poorer performance after frontal injury than in healthy people, with mixed evidence and no one-to-one mapping onto that part of the brain. A failed card sort records a damaged person doing a task that needs many abilities at once.

A computer’s scheduling policy is written by someone else and runs the same way every time. A person’s is partly deliberate, partly habit, and partly set by whoever designed their surroundings, the way the children’s waiting tracked whether the experimenter had kept earlier promises. Over a week of beeper reports, people high in self-control described fewer and weaker temptations, and resisted no more often than anyone else. Beneficial habits accounted for part of the link between self-control and good outcomes in another set of studies.

So the laboratory component keeps dissolving under isolation, while behavioural self-control, rated over years in Dunedin, predicted health, money and convictions at thirty-two. Executive function behaves less like a component and more like a policy for arranging behaviour around a goal, written partly in the head and partly in the surroundings. Real differences in ability and real frontal injuries stay in the picture; the arrangement takes its share of the explanation. In Mischel’s experiment, covering the prize bought children with no instruction at all nearly thirteen minutes.

Laboratory executive tasks and questionnaires about everyday executive function barely agree. Across twenty studies the typical correlation between them was .19. In a larger project that gave 522 people both kinds of self-regulation measure, the average was close to zero. Someone who follows an examiner’s changing rules quickly can still fail to organise a week, and someone who organises a week well can be ordinary on a timed task.

The laboratory supplies the goal, the instructions and the moment to act. An examiner has already decided what comes next and keeps the participant inside the task, which removes the part that tends to go wrong at home. A questionnaire asks whether goals hold up across ordinary days.

Questionnaires also kept their ranking of people better when the same participants came back about 111 days later. Tasks did worse, and scores built by subtracting one condition from another did worst. In the same project, surveys predicted about a tenth of the variation in everyday outcomes such as smoking, drinking, income and life milestones. Tasks predicted about one per cent. The questionnaire asks how days actually go, habits and arrangements included, and that’s where the prediction sat.

The outcomes were self-reported too, which flatters questionnaires: two answers from the same person can agree for reasons that have nothing to do with executive control. Combining the tasks into reliable scores still left their prediction near one per cent.

Three patients with frontal injuries made the gap concrete in a classic study. Two did well on standard frontal tests. All three struggled badly with open-ended errands lasting fifteen to thirty minutes, and typically spent too long on individual tasks. Nobody told them when to stop one thing and start the next, which is the job the examiner does in every standard test.

Executive performance reaches adult levels well before the familiar claim that the brain finishes at twenty-five. A study of 10,766 people aged eight to thirty-five found rapid development between ten and fifteen and adult levels by eighteen to twenty. A frequently cited 2004 brain-scanning study followed children only from four to twenty-one. It showed parts of the prefrontal cortex among the last regions to change, and couldn’t have measured a finishing point at twenty-five. Adult-level scores on a set of tasks still say little about judgement in every setting.

Much of what looks like executive ageing is general slowing. In a review of twenty Stroop studies, older adults were about 1.9 times slower across every condition, and that slowing accounted for their extra interference. Other executive measures likewise had few links with age of their own once reasoning and speed were counted. One study found older adults worse at stopping a dominant response and better at ignoring distracting interference, two directions a single inhibition score would average away.

Keeping two sets of rules ready is the exception. Reviews found an age cost beyond general slowing for global switching and for doing two tasks at once, and none for changing rules from one trial to the next.

Phineas Gage’s recovery fits the arrangement account better than the ruin story. After a railway accident in 1848 drove an iron rod through his head, his friends said he was no longer Gage. He later spent roughly seven years driving coaches in Chile, and a historical reconstruction argues that the job’s fixed routine helped rehabilitate him. It’s a reconstruction from scattered records, and the coach-driving years still sit badly with the story that his early personality change decided his future.

Childhood self-control also reaches into biological ageing. In Dunedin, children rated across their first decade aged more slowly in their bodies by forty-five, and that held after adjusting for childhood intelligence and family class. The links with brain ageing faded under the same adjustment. It’s an observational result about midlife ageing and readiness for old age, and the study is too young to say anything about lifespan.

Stress and illness move scores on shorter timescales. Acute stress impaired working memory and flexibility while possibly sharpening the ability to stop an action, and depression brought broader impairments that grew with severity. Attention deficit hyperactivity disorder (ADHD) carried medium-sized average weaknesses across eighty-three studies, which the review’s authors judged neither necessary nor sufficient to cause all cases.

The Trail Making TestTrail Making TestA pencil-and-paper task that times how quickly a person connects scattered targets in order. Its second part alternates numbers and letters, adding a changing rule to the tracing. is the grip dynamometer of executive function: cheap, timed, and published with age norms. Part A connects the numbers one to twenty-five in order. Part B alternates numbers and letters, 1-A-2-B-3-C, up to thirteen. A pencil, paper and stopwatch are enough, and free printable versions exist.

Part B averages about a minute for adults under fifty-five. In a norming study of 911 adults, mean times were about forty-nine seconds at eighteen to twenty-four, fifty-one at twenty-five to thirty-four, fifty-eight at thirty-five to forty-four and sixty-four at forty-five to fifty-four. Education mattered mainly after fifty-five, and the youngest group was all university students. These are reference points from community volunteers; one slow attempt diagnoses nothing.

Trail B leans mostly on working memory and secondarily on switching. Subtracting the A time strips out much of the visual search and tracing, and the study that showed this, with forty-one people, read the difference as a fairly pure index of switching. Retest studies report B as reliable, and practice still makes the second attempt faster.

Home timing differs from the clinic, where an examiner points out each mistake while the clock keeps running. Write down both times, any errors you notice and the conditions. Compare later attempts with your own earlier ones, allowing for having seen the task before. Then set the trend beside what happened to your appointments and deadlines over the same months, which is the part the questionnaires were measuring.

Training one executive component improves that component and, in pooled studies of children, little else. Broader brain-training reviews found no reliable gain in intelligence, school skills or everyday cognition.

Willpower as fuel has fared badly in large replications. The original experiment had people resist chocolate and eat radishes, then timed how long they persisted on an unsolvable puzzle; the radish-eaters gave up sooner. A test across twenty-three laboratories found almost no effect, and so did a later one across thirty-six, led by Kathleen Vohs, one of the original programme’s own researchers. A third, across twelve laboratories with a different pair of tasks, found a small effect, a tenth of a standard deviation. The glucose explanation, that restraint burns blood sugar, found no support in a separate meta-analysis. Feeling drained after restraint is real; a large, general fuel effect isn’t.

Exercise is the lever this series shares with physical longevity, and its cognitive effect is disputed. In healthy people, a small overall gain fell to almost nothing after correcting for publication bias, and a 2026 correction to that review left the conclusion standing. A broader review that included clinical groups found a small executive benefit, larger in children and people with ADHD, and single bouts gave adults aged eighteen to forty-five small gains in inhibition and working memory, mostly in speed. Reviews rated lowest in confidence reported the largest effects.

Methylphenidate improved inhibition more clearly in ADHD than stimulants did in healthy adults. Mindfulness against active comparisons was small and inconsistent. The adult bilingual advantage disappeared after publication bias was accounted for, which leaves language learning with all its other reasons intact.

An implementation intentionimplementation intentionAn if-then plan that links a recognisable situation to a specific action. It decides the response ahead of time, so the plan is ready when the situation arrives. gives a goal its trigger in advance. Across ninety-four tests, the typical person with one reached their goals more reliably than about three in four people without one. The outcome there is goals reached, a different measure from an executive test score. A deadline rule might read, “If a request conflicts with the agreed deadline, I’ll defer it until the current task is finished.” “I’ll stay focused” leaves that decision open until the request arrives.

An external pager supplied a missing trigger for 143 people with memory and planning problems who completed a sixteen-week trial, most after head injury or stroke. In the randomised crossover trial of NeuroPage, completion of everyday target activities, such as taking medication, keeping appointments and self-care, rose from under half to about three in four. The group still waiting for its pager stayed under half, and the first group still completed about three in five after handing its pager back.

Every lever that reliably moved daily completion here worked on the arrangement. An if-then plan settles the response before the situation arrives, and a pager supplies the start the person can’t. Neither asks anyone to try harder, and the pager trial ran in people with substantial impairments.

Cognitive offloadingCognitive offloadingChanging a task by putting some of its demands into the environment, for example setting a reminder instead of relying on unaided recall. It reduces the work the person must carry internally. works for intentions. In online experiments, people set more reminders as the memory load rose, and reminders improved how many of the instructed actions they carried out. Remembering an appointment unaided earns nothing extra if a prompt gets the person there.

People misjudge how much help they need. Older adults in a preregistered reminder study were overconfident about their memory and set too few reminders to cover their poorer performance. Having the tool didn’t settle when to use it, and a rule decided in advance puts the reminder in place before confidence decides it’s unnecessary.

The evidence that artificial intelligence (AI) damages thinking is much weaker than the evidence that reminders help. “Your Brain on ChatGPT” is a preprint, still without a peer-reviewed version, whose crossover session had eighteen people split between two arms. Its language-model writers showed lower measured brain connectivity and remembered less of their own essays, and lower activity while a tool does the work is what you’d expect. The small crossover also reported weaker connectivity when former model users later wrote unaided, a result that needs replicating before it means anything. A survey and interview study of 666 people linked heavier AI use with lower critical-thinking scores, but it was cross-sectional, so people who already thought less critically may simply use AI more. Neither shows that delegating execution degrades executive function.

Someone sets the priorities an agent works under. A feed or an agent whose priorities the person didn’t set can supply the next goal along with the means to reach it, and a longer list of finished tasks says nothing about whether they were the right ones.

My agents read standing rules files, written in advance, before they start a task. Those files are a scheduling policy, written down.

Questionnaires predicted everyday life where laboratory tasks barely did, and they’re the measure that asks how a person’s days actually go. The pager and reminder studies put the prompt and some of the remembering outside the head, while the people still did the tasks; handing the tasks themselves to an agent is the next step, and it’s untested. None of it decides which goals deserve the day. Let the agent take the context switches, and keep writing the policy that decides when they happen.