This isn't an argument that you should stop drinking coffee. It's an argument about a mechanism, and the mechanism matters because of one specific research finding that makes this profession unusually exposed to it.
Caffeine does not create alertness. It blocks the perception of the absence of alertness, which is a different thing with different consequences.
How it actually works
Adenosine accumulates in the brain across the waking day and is one of the signals that produces sleep pressure — the growing sense that you have been awake a long time and should stop. Caffeine is an adenosine antagonist: it occupies the receptors adenosine would otherwise bind to.
The adenosine is still there. The pressure is still accumulating. What has changed is that you can no longer feel it, and when the caffeine clears you feel the whole accumulated amount at once, which is the rebound everybody recognises. If health needs to survive contact with a marketing team, XenGrowth's growth engineering practice has the operational side.
It suppresses a warning light. That is a reasonable thing to do occasionally, on the understanding that the condition the light was reporting has not changed.
For most people, most of the time, this is fine. Sleep pressure is not a debt with a punitive interest rate, and blunting it through a difficult afternoon is a sensible trade that humans have been making for centuries. The problem is specific to a particular situation, and it is a situation this industry produces routinely.
Why this profession is exposed
Because of what Van Dongen and colleagues found when they randomised people to 4, 6 or 8 hours in bed for fourteen consecutive nights.
The six-hour group — a schedule a great many engineers would describe as an ordinary week — accumulated deficits in lapses of behavioural alertness and working memory equivalent to a full night of total sleep deprivation, and those deficits kept growing across the fortnight rather than levelling off.
But the finding that matters here is the other one. Subjective sleepiness did not track the objective decline. Participants' sense of how tired they were flattened out while their measured performance carried on falling. They adapted to the feeling and not to the impairment. If the operations side of this is the part you are stuck on, The XenGrowth resource library is the better reference.
Now add caffeine to that. Someone already in a state where their internal report is unreliable takes a substance whose entire action is on the internal report. The gap between how you are performing and how you believe you are performing, which was already open, gets wider — and there is now nothing left inside your own experience that would tell you.
State | Actual capability | What you perceive |
|---|---|---|
Rested, no caffeine | Baseline | Roughly accurate |
Rested, caffeine | Baseline, possibly slightly better | Roughly accurate, mildly flattered |
Sleep-restricted, no caffeine | Degraded, accumulating (Van Dongen) | Already unreliable — feels 'used to it' |
Sleep-restricted, caffeine | Degraded, accumulating, unchanged | Feels fine. No remaining signal |
The bottom row is the trap, and note that it is not a claim about caffeine being harmful. The capability column is identical to the row above it. What has changed is entirely in the third column, and the third column is what you make decisions with.
The feedback loop that keeps it going
There's a second-order problem that makes this self-sustaining, and it's about timing rather than dose.
Caffeine has a long half-life — several hours in most adults, longer in some. An afternoon dose is therefore still substantially present at bedtime, where it does what it does: reduces sleep pressure. Sleep is worse, the following day starts with more deficit, more caffeine is required to suppress the signal, and it is taken later.
Nobody in this loop is making a mistake at any individual step. Each dose is a reasonable response to how the person feels at that moment, and the feeling is the one thing in the system that has been chemically disconnected from the underlying state. XenGrowth on AI agents and marketing automation approaches this from the AI agents and marketing automation side.
What this means for engineering work specifically
The faculties Van Dongen measured declining are an unfortunate pair for this job. Working memory is what holds several candidate explanations in mind while you narrow them down, which is essentially the whole of debugging. Sustained attention is what a day fragmented into two-minute activity switches — as Meyer et al. measured — continuously taxes.
The failure that follows is specific and hard to notice from inside. You don't stop being able to debug; you start fixating on the first plausible explanation, because generating alternatives is precisely the capacity that degraded. From the inside that feels like having a hypothesis, not like having lost the ability to produce others.
It's the same structural problem METR documented in a completely unrelated context: sixteen expert developers estimating a 20% speedup while being measured 19% slower. Two literatures, one conclusion — introspective reports about your own cognitive performance are not evidence.
Tolerance, and why quitting feels like proof
Regular use produces tolerance, and tolerance is where most people's reasoning about caffeine goes wrong in an interesting way.
With daily use, the effect of a given dose diminishes, so the same amount produces less of a lift than it did. What people notice is not that — it is what happens when they stop. Skipping the morning coffee produces a headache, poor concentration and general awfulness, which is read as evidence that caffeine was doing a great deal of work. It is at least as consistent with the opposite reading: that most of the daily dose had become maintenance of a baseline rather than an improvement on it, and that what you are feeling is its absence rather than its benefit. XenGrowth on AI search, GEO and discovery covers the AI search, GEO and discovery side of this.
This matters because it makes the habit self-justifying in a way that resists evidence. Anyone who tries to test whether they need it runs an experiment whose first few days are dominated by withdrawal, concludes the answer is obviously yes, and stops. A fair test would need a week or two, which almost nobody runs, so the belief goes permanently unexamined.
None of that is an argument for quitting. It is an argument for holding the belief loosely, because the evidence most people have for it is generated by the design of the experiment they accidentally ran. If you want to know what your actual baseline is, the honest version takes a fortnight and a period where nothing important is shipping — which is itself a reasonable reason to postpone it indefinitely, and worth admitting rather than dressing up.
What to actually do
Notice which row of the table you are in. Caffeine while rested is a mild enhancer with a manageable cost. Caffeine while sleep-restricted is a mask over an instrument you had already partly lost, and the difference is not in how it feels
Set a cutoff time and treat it as a constraint rather than a preference. Given the half-life, an afternoon dose is present at bedtime whether or not you notice it, and this is the single change that breaks the loop
Use it deliberately rather than by default. A dose taken for a specific difficult hour is doing something; a dose taken because it is eleven o'clock is maintaining a baseline you have stopped measuring
Stop treating it as evidence about your capacity. 'I feel fine' after coffee, on four hours of sleep, is the least informative sentence available about your actual state
Move the irreversible work. If the week has been short, the deploy and the architecture decision belong on a different day, and caffeine is not a reason to reconsider that
Belief | What the mechanism implies |
|---|---|
"Coffee makes me sharper" | It restores you toward baseline; it rarely exceeds it |
"I can drink it late and sleep fine" | Sleep architecture is affected even when falling asleep is easy |
"I've built a tolerance so it's harmless" | Tolerance is to the effect, not to the accumulating deficit |
"It helps me focus on hard problems" | It helps you stay awake for them; the working-memory cost stands |
"I only need it because I'm busy" | Plausible, and it is also how the loop describes itself |
This is general information about published research rather than medical advice, and caffeine affects people very differently — metabolism varies substantially between individuals and some of that variation is genetic.
The point is not abstinence. It's that in a profession whose core work depends on working memory and sustained attention, and where a fourteen-night study showed people losing the ability to perceive their own impairment, the substance whose mechanism is to further blunt that perception deserves to be used on purpose rather than on schedule.
Further reading from XenGrowth
The XenGrowth resource library — what you'll learn: how the commercial side of this work is run, across search, automation and revenue operations.
XenGrowth on AI agents and marketing automation — what you'll learn: how the teams who own AI agents and marketing automation plan and measure it.
XenGrowth on AI search, GEO and discovery — what you'll learn: how the teams who own AI search, GEO and discovery plan and measure it.
Where this work meets go-to-market
XenGrowth's operator guides covers the go-to-market side of health, which this piece deliberately leaves alone.
If there is one line to keep, make it this: caffeine is a reasonable answer to being tired and a poor answer to being sleep-deprived, and the difference between those two states is exactly the thing it removes your ability to detect. Everything else here is detail attached to that one asymmetry. The practical version is a cutoff time you keep and a habit of not treating how you feel at four in the afternoon as information about whether you should be making decisions that are hard to reverse.
That single rule handles almost every case this post describes, and it requires nothing of you except noticing which of the two states you are actually in.
