The only thing that stops me is a usage bar

My weekly allowance on one of the AI tools I use reset on Tuesday. By Wednesday afternoon I’d spent 98 per cent of it. I screenshotted the bar and sent it to my best mate, who is also my head of technology. He replied: “You’re an absolute monster with it bro.”

I put the phone down feeling great about myself. That’s the bit worth writing down.

How the week actually goes

I work with two agents most days, Claude and Codex. One builds while the other reviews, or they take different jobs and come back at different times. Neither of them gets tired. Neither of them has a Tuesday.

One of them does tell me to stop. Claude will finish a job at an unreasonable hour and put a line at the end of it like:

That’s a good stopping point. You’ve been at this a while, and the next bit will be easier with fresh eyes in the morning.

It infuriates me. I swear at it. I have sworn out loud at a language model for suggesting I go to bed, which says more about me than it does about the model. Then I ask it for the next thing, and it does the next thing, because that’s the one instruction it never argues with.

So the machine holds the stopping rule and has no standing to enforce it, and I have the standing and don’t use it.

For most of my working life the day ended because of other people. Someone else had to do the next bit, or the office emptied, or I got tired enough that the work went obviously bad and I could see it going bad in front of me. Those signals are harder to read now. The queue refills faster than I can clear it, and the drafts keep coming when I’m tired. Whether I’m still checking them properly is another question.

The only thing that stops me is a bar on a billing page, filling up on a schedule set by a company that would like me to buy more. As stopping rules go, it’s crap. It measures how much of my allowance I’ve used. It knows nothing about me.

The honest part

My brain ran at a hundred miles an hour before any of this arrived. I’m neurodivergent, I’ve always had more going on in my head than I could get out of it, and the constraint on my working life was never ideas, it was throughput. These tools took the throughput constraint away. Apparently the bottleneck was also the brakes.

The loop is the addictive bit for me. Ask, wait, get something back, react. It’s quick, I don’t know what I’ll get, and I keep going back for more.

At the start I couldn’t keep up with it. One model, one thread, and I’d send a prompt, read what came back properly, argue with it, and still be behind. The bottleneck was me and I knew it.

Now I run two of them in unison, Claude on Fable and Codex on GPT-6 Astra, several threads at a time. My most common feeling while I wait is irritation. I sit there stressed that they aren’t keeping up with me. Even that wait feels too long, so I switch to the other thread to fill the gap, which means there is no longer any point in the day where I’m waiting for anything.

What I can tell you is that the setup which amazed me at the start now reads as slow, and nothing about the setup got worse.

Occupational psychology gives me a way to think about this. The job demands-resources model distinguishes what work asks of you from what helps you do it. High demands can contribute to exhaustion; resources can support engagement and buffer the strain.¹ An agent arrives looking like a pure resource. It adds capacity, it does the work, it is sold to me as the thing that takes load off. I’ve spent all of it on more demand, and I’m the one setting the level.

I’m going to use the word addictive and then not reach for a dopamine study to support it, because in July an audit of something I wrote caught me citing a paper that didn’t say what I claimed it said. A confident paragraph about dopamine would do very well on LinkedIn and I’ve no business writing one. What I can report is behaviour. I work myself to the bone. I check what came back before I’ve eaten. I have never once finished a day because the work ran out.

What the evidence separates people by

Burn-out has a definition, and it’s narrower than the way most of us use the word. The World Health Organization describes it as a syndrome resulting from chronic workplace stress that has not been successfully managed, with three dimensions: exhaustion, mental distance or cynicism about the job, and reduced professional efficacy. It sits in the classification as an occupational phenomenon rather than a medical condition.² Nothing in that definition counts hours.

In 2008, Schaufeli, Taris and van Rhenen surveyed 587 Dutch telecom managers to test whether workaholism, burnout and work engagement are three of a kind or three different things. The results supported three distinct but related constructs. The managers scoring high on engagement had good mental health and smooth social functioning, and they also worked long hours. The managers scoring high on workaholism worked hard too, and had a similar pattern of health and social difficulties to managers scoring high on burnout. Workaholism also went with poor job control and little supervisory support, which the authors flagged as an unexpected finding.³

Both engagement and workaholism were associated with long hours. Poorer job control also went with workaholism; that doesn’t tell me what happens when I’m the one choosing the work. That’s one cross-sectional survey of one narrow group of managers in one country, with mostly small to moderate effects, and the authors name the cross-sectional design as their most important limitation. I wouldn’t hang a policy on it.

A much bigger survey points the same way. Langseth-Eide sampled 12,170 employees at Norwegian universities and university colleges and tested both concepts inside the demands-resources model. Job demands were associated with higher workaholism scores and job resources with higher engagement. Workaholism went with worse perceived work-related health, and engagement with better. Both were associated with overtime, more strongly for workaholism.⁴ Cross-sectional again, self-report again, one country and largely one sector, so a large snapshot rather than a demonstration of cause.

It’s still a better question than the one I’d been asking myself, which was whether I’d done too many hours this week. That leaves me with two questions of my own: could I stop, and would stopping feel fine? First one, probably yes. Second one, I don’t know, because I don’t test it often enough to have an answer.

Recovery is the thing the agents are pointed at

The evidence that has held up best here is about detachment, meaning away from work in your head as well as out of the building. A 2017 meta-analysis pooled 86 publications, 91 independent samples and 38,124 employees. Detachment was associated with less exhaustion, higher life satisfaction, better sleep and lower physical discomfort, with small to medium effects. Its relationships with physiological stress indicators were not significant, and it was negatively associated with contextual performance and creativity.⁵ Mostly self-report, mostly cross-sectional, so associations rather than causes.

An always-running agent makes it harder for me to detach. A job that finishes at two in the morning is a reason to look at your phone at two in the morning. I’ve argued elsewhere that agents should hold their output until the person’s next working period, and I still think that. I also know I’d override the setting inside a week.

The evidence assembled for the review did not identify a study pairing the timing and intensity of heavy, self-directed agent use with recovery, sleep or burnout. That search has limits, which the review sets out, so read it as a gap this search couldn’t fill rather than proof there’s nothing there.

So I checked it properly. It’s one of the current midnight obsessions, which I appreciate is the whole problem in one sentence. I’ve put the evidence review alongside this note. It covers recovery and sleep, problematic use, job control and neurodivergence, with the sources and their limits attached. If the evidence changes my account, the correction gets a date. It’ll be here for anyone that way inclined.

What I’d want instead

None of this is an argument for doing less. I’m not going to do less. The work is good at the moment and I want to do it.

What I’d want is something in my week that measures me rather than my usage. A bar that fills up tells me how much of my allowance I’ve used. A mate calling me a monster is applause. The allowance resets on Tuesday, the queue will be deeper for the wait, and I’ll be pleased about that too.

Bakker, A. B. and Demerouti, E. (2017). Job demands-resources theory: taking stock and looking forward. Journal of Occupational Health Psychology, 22(3), 273-285. doi:10.1037/ocp0000056. A review of the model and the evidence behind it rather than a single study. Full text via Erasmus University Rotterdam, p. 274, propositions 1 to 3. Relevant full-text sections checked 16 September 2026. ↩
World Health Organization (2019). Burn-out an “occupational phenomenon”: International Classification of Diseases. News release, 28 May 2019, who.int. The release places it in the chapter on factors influencing health status or contact with health services, and states that it is not classified as a medical condition. Read 16 September 2026. ↩
Schaufeli, W. B., Taris, T. W. and van Rhenen, W. (2008). Workaholism, burnout, and work engagement: three of a kind or three different kinds of employee well-being? Applied Psychology: An International Review, 57(2), 173-203. doi:10.1111/j.1464-0597.2007.00285.x. 587 telecom managers in the Netherlands, 69 per cent response rate, cross-sectional. The authors describe the sample as homogeneous and the cross-sectional design as the most important limitation. ↩
Langseth-Eide, B. (2019). It’s been a hard day’s night and I’ve been working like a dog: workaholism and work engagement in the JD-R model. Frontiers in Psychology, 10, 1444. doi:10.3389/fpsyg.2019.01444. 12,170 employees at Norwegian universities and university colleges, cross-sectional, self-report, structural equation modelling. ↩
Wendsche, J. and Lohmann-Haislah, A. (2017). A meta-analysis on antecedents and outcomes of detachment from work. Frontiers in Psychology, 7, 2072. doi:10.3389/fpsyg.2016.02072. 86 publications, 91 independent samples, 38,124 employees. ↩