The Twentieth Watt
On the depletable collaborator, and the reservoir that empties while you think it works
There are cafés that serve coffee, and there are cafés that serve as instruments for measuring what you have spent.
Le Bon Mot was quiet in the way it becomes quiet after nine in the evening. Not empty, but contracted, compressed somehow, defragged, the shelves pulling inward as if the room itself were settling into a denser configuration for the night. The brass clock behind the counter showed 9:07, which meant the actual time was 9:11, which meant that Madame Beauregard had already begun her end-of-evening ritual of wiping down surfaces that were already clean and adjusting the position of books whose spines had shifted fractionally since the last time she adjusted them.
Sophie was asleep on the third shelf from the floor, curled into a space between a volume of Epictetus and a systems engineering manual that had not been there that morning. Her breathing had the slow regularity of a creature that has never once questioned whether it has done enough today.
The bell rang. The door opened. And the person who entered was incandescent.
Not literally, of course. Le Bon Mot had unusual properties, but spontaneous bioluminescence was not among them. The incandescence was professional. The kind that radiates from someone who has had the best working day of their life and needs, urgently, to tell another human being about it.
“You,” the person said, pointing at Madame Beauregard with the hand that was not holding a laptop bag, “are looking at someone who just shipped three features in a single day.”
Madame Beauregard set down the glass she was polishing. “Congratulations.”
“Three features. Six agents. Every PR merged green. The harness caught everything that needed catching. The specs were clean, the decomposition was tight, and I conducted the entire thing like—” They searched for the word. “Like a composer.”
“Sit,” Madame Beauregard said. “I’ll make you something.”
The person’s name was Jules. They sat at the counter with the satisfied exhaustion of someone who has crossed a finish line they set for themselves. Their laptop emerged from the bag. The screen showed a dashboard: three pull requests, three green checkmarks, three merge timestamps spanning from 10:42am to 7:18pm.
“Look at that,” Jules said. “I have never had a day like this.”
Madame Beauregard placed a cortado on the counter. The espresso machine, which had been silent all evening, produced a sound that might have been steam releasing, or might have been Latin expressing mild scepticism.
“Tell me about the morning,” Madame Beauregard said.
Jules did not need to be asked twice.
“I started with the authentication refactor. The spec was already solid from last week — I’d written the acceptance scenarios Friday afternoon, and Monday morning I just needed to decompose it into agent tasks. Three agents: token validation, session management rewrite, migration tests. Each one got a context file with the architectural constraints, the existing API contract, and the specific scenarios from the spec.”
They took a sip of their drink. “And I reviewed every output like a code review from a junior engineer I respected but didn’t yet trust. Line by line. I ran the edge cases in my head before I ran them in the suite. The token agent generated a validation function that passed all the specified tests but would have silently accepted expired tokens with a clock skew greater than thirty seconds. I caught it because I had a model in my head of how the function should fail, and the generated version was too permissive.”
Madame Beauregard nodded. “You had a theory of the change.”
“Every approval was a decision. I could have explained each one — why this test mattered, why that design choice was right for our domain, why the agent’s suggestion to use a different data structure was clever but wrong for our access patterns. The whole thing took until about eleven. First PR, merged green.”
Sophie shifted on her shelf. The Epictetus moved a quarter-inch to the left, as if making room for something that had not yet arrived.
“And then?” Madame Beauregard asked.
“Then the notification service. That one was bigger — two agents on the core event pipeline, one on the delivery abstraction, one on the retry logic. I had the spec decomposed by eleven-thirty. The agents were running by noon.”
Jules described it with less granularity than the morning. The event pipeline had been the interesting part — a publish-subscribe pattern replacing the old polling mechanism. The retry logic had edge cases around idempotency that the agent had handled well.
“Green by about two-thirty. Second PR, merged.”
“And the third?”
Jules paused. Not for long, barely a breath. But the rhythm changed.
“The third was the dashboard analytics. Aggregate query layer on top of the event data from the notification service. I had the agents set up by three. Specs were ready — I’d prepped them over lunch.” A pause. “Four agents this time.”
Madame Beauregard waited.
“It came together fast. The caching strategy was elegant. The API was clean. Tests all green.” Jules gestured at the screen. “Merged at 7:18.”
“Tell me about the final review,” Madame Beauregard said.
“What about it?”
“The last PR. The analytics dashboard. When you approved it. What specifically did you verify?”
Jules opened their mouth. Closed it. Opened it again.
“The tests were green,” they said. “All the scenarios from the spec passed. The linter was clean. The harness ran without warnings.”
“Yes. The automated checks. What did you verify? You, personally, with your eyes on the diff.”
The silence that followed was someone reaching for a memory and finding it less substantial than expected.
“I read the diff,” Jules said. “I remember reading it.”
“Do you remember what you thought about the caching strategy? You said it was elegant. What made it elegant?”
Jules stared at the green checkmark on the screen. “It used a… There was a TTL-based invalidation on the...” They trailed off. “It was a well-structured cache. The agent chose a good pattern.”
“Which pattern?”
“I’d have to look at the code.”
Madame Beauregard let the statement settle. She did not press. She picked up the glass she had been polishing and resumed polishing it, though it had been clean since before Jules arrived.
“The notification service,” Madame Beauregard said. “The second feature. Tell me about the retry logic.”
Jules recovered some footing here. “Right, the retry logic. That was tricky — we needed idempotency guarantees because the downstream consumers aren’t all built to handle duplicates. The agent generated a deduplication mechanism using...” They paused. “Using a hash of the event payload as the idempotency key.”
“Was that the right choice?”
“Yes, because—” Jules stopped. “Actually. I remember thinking there was a question about whether payload hashing was sufficient, or whether we needed a separate idempotency token from the producer. There was a design choice there.” A longer pause. “I think the agent made the case for payload hashing in a comment, and I... agreed with it.”
“Did you evaluate the case, or did you agree with the conclusion?”
Jules looked at the dashboard. Three green checkmarks. The same shade of green at 10:42am and 7:18pm.
“I don’t know,” Jules said quietly. “I genuinely don’t know if I evaluated it or just... read it and moved on.”
“And the morning’s work? The authentication refactor. The token validation that was too permissive?”
Jules straightened. “I caught that because I had run the failure scenarios in my head before I looked at the code. I knew what shape the rejection should take, and the generated version was the wrong shape. I could draw you a diagram of the token lifecycle right now.”
“I believe you.” Madame Beauregard set down the glass. “Three features. Three different levels of presence from the same reviewer.”
Jules sat with this. The brass clock ticked. Sophie’s ear twitched. A book on the second shelf — one that Jules could have sworn was a dictionary — now appeared to be titled On the Shortness of Sessions.
“I didn’t get worse at my job today,” Jules said. “I’m the same engineer at seven pm as I was at nine am.”
Madame Beauregard set down the glass. “You are the same engineer. You were not the same reviewer.”
She reached behind the counter and produced a piece of chalk. On the chalkboard, which had previously displayed the evening specials (a cortado, a lungo, and something called “The Cartesian Doubt” which was espresso served with the question but is it really espresso?), she wrote a number.
20
“The brain,” she said, “runs on approximately twenty watts. You know this.”
“Everyone knows this. Less than a light bulb.”
“Less than a light bulb. And with those twenty watts, it does everything — language, pattern recognition, planning, emotional regulation, motor control, the management of a self that persists across time. But here is what the watt comparison obscures: those twenty watts are not a fixed supply. They are a reservoir.”
She drew a shape on the board — a vessel, wide at the top, narrow at the base.
“The reservoir fills during recovery — sleep, rest, disconnection. And it drains with use. Not uniformly. Decision-making drains it. Context-switching drains it faster. Sustained verification — holding a model in your mind and comparing output against that model — drains it fastest of all. Because verification requires you to simultaneously hold what is, what should be, and the gap between them.”
Jules was listening with the attention of someone hearing a description of something they had felt but not named.
“Your agents ran at the same quality all day,” Madame Beauregard said. “Their output at three o’clock was as sharp as their output at nine. No degradation. No fatigue. The machine does not have a reservoir.”
She tapped the drawing.
“But you do. And by the third feature — by hour nine of sustained composition, of context-switching between four parallel agents, of reading diffs and making approval decisions — your reservoir was not what it was in the morning. The morning’s twenty watts had become the evening’s eight.”
“I didn’t feel tired,” Jules said.
“No. You wouldn’t. That is the important thing about cognitive depletion: it degrades your ability to detect it too. The faculty you would use to judge whether you are still sharp — metacognition, self-monitoring — draws from the same reservoir as everything else. When the reservoir is low, you lose the instrument that would tell you it’s low.”
The espresso machine hissed. This time, it was unmistakably Latin:
Nemo iudex in causa sua.
No one is a good judge in their own case.
“Your harness,” Madame Beauregard continued, “enforces architectural constraints. It runs mutation tests. It checks for convention drift. It verifies that the agents’ output conforms to your specifications.”
“That’s why I trust it.”
“And what does it check about you?”
Jules blinked.
“No hook verifies whether the reviewer was cognitively present when they approved the PR. No CI gate measures whether you built a theory of the change or merely watched it pass.”
She let that land.
“The harness verifies the output. It cannot verify the verifier. And you are the verifier. The judgment that says yes, this belongs in our system — that runs on a reservoir that was near empty by the time you approved your third feature.”
Jules looked at the three green checkmarks. Same colour as ten minutes ago. They no longer communicated the same thing.
“So I should have stopped.”
“That is not quite the question. When should you have decided to stop?”
“When I got tired.”
“When you noticed you were tired? We just established that the instrument for noticing is the instrument that degrades. If you wait until you feel you should stop, you have already passed the point where the decision was trustworthy.”
Madame Beauregard wrote on the chalkboard, below the vessel:
The decision to stop must be made before the session begins — when the judgment that makes the decision is still the kind you would trust.
“Epictetus told us that the only thing fully within our control is our own prohairesis — our capacity for reasoned choice. Not the agents’ output. Not the harness’s enforcement. Whether we continue to exercise judgment when the faculty of judgment has been spent, that choice is ours.”
She set down the chalk. “And it is the one choice the framework cannot make for you.”
Sophie opened one eye. She had been awake since approximately the moment Madame Beauregard drew the vessel on the chalkboard, but had seen no reason to make this known. She regarded Jules with the expression French bulldogs reserve for truths acknowledged but not yet absorbed.
Jules looked at the laptop. Three merged PRs. The dashboard showed what any external observer would call an extraordinary day.
“Should I go back and re-review the last two?”
“Now?” Madame Beauregard smiled. “At 9:20 in the evening, after the day you’ve described?”
Jules almost laughed. “Right.”
“Tomorrow morning. When the reservoir is full. That is when you will see what tonight you could not. You may find the work is excellent. You may find decisions you would not have made. But either way, you will be seeing with the eyes that caught the token validation error. Not the eyes that approved the caching strategy you cannot now describe.”
Jules closed the laptop. The same gesture as closing a book you intend to reopen.
The chalkboard had changed again. Where the vessel and the number 20 had been, there was now a single line in a hand that belonged to no one present:
The harness does not sleep because it does not need to. You do, because you must.
Sophie settled her chin back onto her paws. The Epictetus on the shelf beside her had fallen open to a less-quoted line from the Discourses: It is not enough to be willing. You must be able.
The brass clock showed 9:22. Which meant it was 9:26. Which meant that Jules had been at Le Bon Mot for fifteen minutes, and had learned something about the previous twelve hours that the previous twelve hours had not been able to teach them.
Madame Beauregard collected the cortado cup. She placed a slim volume on the counter — no author listed, titled simply Sustainable Pace. Whether it had been on the shelf before tonight was a question Le Bon Mot declined to answer.
Jules stepped out into the evening. The door closed. Sophie slept.
The chalkboard, in the silence that followed, added a postscript:
Viginti watt. Hoc satis est — si scias quando desinas.
Twenty watts. It is enough — if you know when to stop.
What Changes at Each Level
The AI Literacy Framework has six levels that map how seriously the engineer takes the fact that they are a finite biological system collaborating with a tireless one.
At Level 0 (Awareness), depletion is not a concept. A long day is a productive day. If code was approved and tests pass, the work is done.
At Level 1 (Prompting), the developer notices retrospectively that their prompts were sharper in the morning. “I was pretty tired by the end” is said on the way out, not used to redesign the next day.
At Level 2 (Verification), depletion signals register in real time. Rubber-stamping reviews. Skipping the understanding check. Irritation at agent output volume replacing curiosity about content. The developer starts to suspect that the green checkmark at 5pm does not mean what the green checkmark at 9am meant.
At Level 3 (Habitat Engineering), recovery is designed into the workflow. “I will review for ninety minutes, then stop” replaces “I will stop when this feature is done.” The developer treats their own cognitive state as a resource to be managed, not an obstacle to be overcome.
At Level 4 (Spec Driven), depletion awareness becomes a team practice. Spec decomposition accounts for human energy, not just logical modularity. A reviewer who has been context-switching across six agents all morning is not the same reviewer after lunch.
At Level 5 (Sovereign), sustainability is an organisational design principle. A team understands that scaling agent count without scaling human recovery is not scaling — it is accelerating debt.
The Depletion Check
Every ninety minutes, ask yourself: Am I still sharp enough to trust my own judgment?
This must be time-based, not judgment-based. You cannot rely on “I’ll stop when I feel tired” any more than you can rely on a thermometer that melts at the temperature you need to measure. The check must be external — a timer, a calendar block, a ritual that fires whether you think you need it or not.
Observable signals that the answer is no:
You approved a review but cannot explain why each change is correct
You skipped the understanding check — you read the diff but did not build a mental model of the change
You feel irritation at the volume of agent output rather than curiosity about its content
Your default has become “looks fine” rather than “let me trace this”
You are starting new agent tasks rather than verifying output from the last batch
Your prompts have become vague or repetitive — copy-pasting context rather than crafting it
Exercise: The 90-Minute Reservoir Check. Set a recurring timer for every 90 minutes of focused AI collaboration. When it fires, stop. Do not finish the current task first. Answer three questions:
(1) What was the last review decision I made, and can I explain the reasoning right now?
(2) Am I starting new agent work to avoid reviewing completed work?
(3) If someone asked me to justify my last three approvals to a colleague, would I be confident in every one?
If you hesitate on any answer, the session is over. Not paused — over. The next session begins after genuine recovery: movement, disconnection, sleep. The reservoir refills on its own schedule, not yours.
Some things to Avoid
The Marathon Session. Jules’s day. Genuinely excellent work in the morning, adequate work after lunch, hollow approval by evening — all three feeling, from the inside, like the same quality of engagement. The marathon session is dangerous because it produces volume. Three features shipped. A full dashboard of green. The evidence of productivity is overwhelming. The evidence of degradation is invisible until someone asks you to describe the caching strategy.
The Always-On Orchestrator. Multiple parallel agent streams, continuous context-switching between them. Each switch carries attention residue — part of your mind stays on the previous task. One parallel stream costs attention. Two costs more than twice as much. Four costs more than four times. The compound tax is multiplicative, not additive. The orchestrator who runs six agents all day is not composing. They are haemorrhaging cognitive capacity while producing impressive activity metrics.
The “Just One More” Trap. The decision to review one more PR, start one more agent task. This decision is always made from the depleted state it should have prevented. The corrective is the session boundary: the decision to stop is made before the session begins, when the judgment making the decision is still trustworthy. You do not negotiate session boundaries with your tired self.
Some Further Reading
The Science
Danziger, Levav & Avnaim-Pesso — “Extraneous factors in judicial decisions” (2011). Israeli parole judges granted parole at dramatically different rates depending on how recently they had taken a break. Favourable rulings dropped from ~65% to nearly 0% across a session, then reset after food and rest. The most concrete evidence that decision fatigue is real and consequential.
K. Anders Ericsson — deliberate practice research. Expert performers sustain deep cognitive work for 4-5 hours per day. The ceiling is biological, not motivational. Musicians, chess players, athletes, and writers all converge on the same limit.
Sophie Leroy — “Why is it so hard to do my work?” (2009). Attention residue: the cognitive fragments that persist after switching tasks. Directly relevant to the cost of parallel agent orchestration.
The Stoic Tradition
Epictetus — Discourses and Enchiridion. The governance of your own cognitive state is the one thing the framework cannot do for you. “It is not enough to be willing. You must be able.”