TL;DR Every hobby I ever had carried its own brake — the soldering iron, the drying plaster, the compiler punishing fatigue. Agents removed it: the cost of starting collapsed to one sentence, and now roughly 90% of the ideas that cross my mind become projects, seven days a week. The loop — idea, prompt, variable reward, one more — is the mechanism the gambling literature describes. And since no peer-reviewed study of agentic coding exists yet, this post does the only thing possible: it compares my case against what gambling science already knows. The ADHD link is a sourced, moderate correlation; for the French "HPI" label, research shows no link to addiction — which is good news, actually. Then the two honest questions: the funnel from idea to living product collapses exactly between "repo created" and "used by someone else" — and the value is real, but it is not where the speed promises it is. First readout from my own account: one new personal repository in over three years before agents, fourteen in six and a half months since — at $200 a month to keep the loop running. Guardrails at the end, because the loop has no built-in exit.

It is past 11 p.m. on a Sunday and I have just relaunched an agent. Nobody is waiting for this project. It has no deadline, no user, no ticket. It exists because the idea crossed my mind in the shower this morning, and between the shower and tonight, the only thing separating the idea from a deployed service was one sentence to type. I typed it. It has been like this seven days a week for months, and it is time to write down what this actually is: an addiction. The word is strong, and I will keep it at arm's length from its clinical definition all the way through — but it is mine.

Until now this series looked outward: the process, the job, the team, the ROI, the six-month verdict. This post looks inward. Three days ago I published the professional verdict — 100% agent-written code on Solstice. Here is the personal bill.

The friction was the brake

In 2016 I was flashing ESP8266 boards. An idea followed a fixed path: order the component, wait five days for the package, solder, flash, debug the voltage divider I got wrong, start over next weekend. The house renovation obeys the same law — plaster takes a day to dry no matter how motivated you are. Every passion I have ever had carried its own brake, built into the physics of the thing. The idea had to survive the wait. Most didn't, and that was the system working.

"The cost of an idea fell to zero. The cost of ten ideas did too."

Then the brake disappeared. Today, roughly 90% of the needs that cross my mind get a software answer within the day — my own number, a felt estimate, not a measurement. The forty-plus services running behind Traefik on my homelab box were each, at some point, a Sunday idea. The distance between "it crossed my mind" and "it runs" is now one prompt long. And that distance no longer even has a door to push open: on my Proxmox VMs, agent sessions stay open around the clock, ready to receive a prompt. The room never closes.

TIME BETWEEN THE IDEA AND THE FIRST ARTIFACT From a weekend to twenty minutes Personal orders of magnitude, ten years apart — a felt estimate, not a measurement ESP8266 · 2016 a weekend — order, wait, solder, flash, debug that time filtered the ideas — it was the brake AGENTS · 2026 twenty minutes — one prompt, one agent, a running service nothing filters anymore — not fatigue, not the soldering iron

Here is the part I underestimated for months: the friction didn't just slow me down. It selected. An idea had to survive a week in my head before it earned a weekend. The weak ones died quietly; the good ones fermented, merged with other ideas, came out sharper. That was natural selection for ideas, and I dismantled it with a subscription.

A loop that looks like something else

Watch the cycle honestly and it has four stations. An idea triggers a prompt. The prompt costs almost nothing. The agent comes back with a reward — and the reward is variable: sometimes it nailed it, sometimes the build is green with one test red, sometimes it confidently broke everything. And the almost-win, more than the win, is what makes me type "one more".

THE LOOP · VARIABLE-RATIO REWARD No built-in stopping point THE IDEA THE PROMPT THE REWARD « ONE MORE » the trigger cost: near zero variable — it almost works the near-miss re-arms THE LOOP « machine zone » : time disappears the exit does not exist by default — it has to be built

None of this is a new mechanism. Variable-ratio reinforcement — reward that arrives unpredictably — is the schedule Skinner identified as producing the highest, most extinction-resistant behavior, and it is the engine of every slot machine. The near-miss has primary neuroscience behind it: Clark and colleagues showed in 2009 that almost-winning recruits the same win-related dopaminergic circuitry as winning, and increases the motivation to continue — a small fMRI study, about gambling, and I am mapping it onto coding by analogy, not by evidence. But an agent that returns a green build with one red test is a textbook near-miss, and I feel the pull it predicts.

The deepest description I know is Natasha Schüll's Addiction by Design, fifteen years of fieldwork among machine gamblers. Her "machine zone": a state where time dissolves, where even winning stops being the point — continuing is the point. And her sharpest claim: this state is a co-production between human and machine design, not an individual weakness. Chew on that one at 11 p.m. while a third agent runs.

"Watching an agent work is not flow. It is anticipation."

I want to be careful here, because the confusion is comfortable: this often feels like flow. But Csikszentmihalyi's flow needs your skill actively engaged against a matched challenge — the doing is the reward. Watching a terminal scroll is not challenge-skill balance; it is waiting for a variable payout. Structurally it is closer to the machine zone than to flow. That framing is mine, not a citation — and it is the most uncomfortable sentence in this post.

What does the actual science say about AI overuse? The closest empirical data is the MIT Media Lab / OpenAI randomized study (2025, 981 participants): assigned usage showed no significant effects, but voluntary heavy users showed higher loneliness, emotional dependence and problematic-use markers. Caveats stacked high: it studies emotional chatbot use, not coding agents; the finding that matters is correlational; it is a preprint co-authored by OpenAI. And beyond it, the honest statement is this: as of July 2026 there is no peer-reviewed study of compulsive agentic coding at all. The genre exists only as testimony — Axios ran "They operate like slot machines" this April, Hacker News hosts a thread literally titled "Addicted to Claude Code – Help". I am writing from pre-scientific territory: a sample of one — me — compared against what the gambling literature has already demonstrated. That is the strongest claim available, so it is the one I make.

The ADHD detour — and the HPI question

Why does this hit people like me harder than others? The reader I have in mind generates ten ideas a day and has spent a life watching nine of them die of friction. For that brain, a machine that converts any passing idea into a reward in twenty minutes is not a productivity tool. It is a perfectly fitted key.

And let's say it up front: in my case, that brain carries no clinical name. No diagnosis, and I do not recognize myself in ADHD. I simply have an outsized capacity to invent, a head permanently full of ideas — nothing more. The ADHD detour is here because the link is easy to make and often made, not because it is personal. So let's look at what the literature actually says, at its real weight.

The literature has pieces of this, and they deserve to be quoted at their real strength — no more. Hyperfocus is a real, described phenomenon: complete task absorption that tunes out the world, discussed mostly in ADHD but not exclusive to it, and overlapping with ordinary flow. The reference review (Ashinoff & Abu-Akel, 2021) says plainly that it is under-researched and inconsistently defined — so treat it as an experience many people will recognize, not a diagnostic marker. It is explicitly double-edged: the same absorption that ships a project at 2 a.m. is the one that loses the evening.

ADHD and compulsive tech use: a 2023 meta-analysis (24 studies, ~19,000 participants) finds ADHD traits correlate with problematic internet use at r ≈ 0.36–0.44 — a moderate, cross-sectional correlation on a trait continuum, found in non-clinical adults too. Not "ADHD causes addiction"; direction unresolved. Underneath, Volkow's PET work (2009) supports a reward-pathway difference in ADHD, consistent with a preference for small immediate rewards over larger delayed ones. In other words: a susceptibility to exactly the loop drawn above. A susceptibility, nothing more. And the entrepreneurship literature (Wiklund and colleagues, 2016 — fourteen case studies, qualitative, survivor-biased) describes the pattern idea-rich builders will recognize: impulsivity as the fuel that starts things, hyperfocus as the catalyst that finishes or consumes them.

"And what about HPI?" In France, this conversation naturally raises the question — HPI, haut potentiel intellectuel, is a common label here — and it deserves a measured answer rather than a punchline. High intellectual potential is a psychometric reality: a measured cognitive profile that real people live with every day. What research does not confirm is the link popular culture often draws from it to psychological fragility or addictive behavior: Franck Ramus and Nicolas Gauvrit's work shows that high-IQ people present no more psychological disorders than anyone else — sometimes fewer. The "tormented gifted" figure owes much to a sampling bias: clinicians only meet the gifted people who come to consult. That is good news, actually. For this post's mechanism, it simply means HPI is not the right lead: if you are looking for what makes this loop land so hard on some brains, ADHD traits and novelty-seeking have support in the literature. High potential, as the evidence stands, does not — and pinning it on the label would be unfair to the people who live with it.

And to be unambiguous: I am diagnosing nobody in this post, myself included. The mechanism describes an experience; it does not hand out labels. If the loop above is eating your nights and it hurts, that conversation belongs with a clinician, not a blog.

First honest question: does any of it ship?

So the factory runs seven days a week. What comes out the other end?

FROM IDEA TO LIVING PRODUCT · ILLUSTRATIVE FUNNEL Velocity fills the top, never the bottom Widths are illustrative — no agent-era completion data exists: count your own Ideas Prompts run Repos created Deployed Used by someone else Alive at 6 months 121M new GitHub repos in one year (Octoverse 2025) ~46% of indie projects at $0 (pre-agent data) the gap Between the repo created and the product someone else uses — that is where the loop stops, not at the prompt.

The top of the funnel has exploded, and that part is measured: GitHub logged 121 million new repositories in one year — about 230 a minute — its biggest year ever. But a repository is not a product; forks and experiments dominate. The bottom of the funnel, meanwhile, looks the way it always has. The indie-hacker datasets we have put roughly 46% of projects at zero revenue, with a brutal power law above them. Those numbers predate agents, are self-reported and survivor-biased: keep the shape, not the decimals. Nobody has published completion-rate data for agent-era side projects. The center of this post rests on a funnel nobody has measured — including me, and I'll come back to that.

What we do have is documentation of what happens when the prototype pretends to be a product. The canonical one: in March 2025 a founder announced a SaaS built entirely with Cursor, "zero hand-written code". Two days later he posted "guys, I'm under attack": hardcoded API keys, no auth, paywall bypassed. He shut it down because he could not debug what he had never understood. One anecdote, self-reported — a perfect illustration, not a statistic. The quantified version: a scan of apps built on Lovable found about 10% exposing vulnerable endpoints (303 endpoints across 170 of 1,645 apps). And the Tea app breach — 72,000 images including government IDs — is widely attributed to vibe-coded infrastructure; attribution is commentary, not confirmation, so I quote it as a security-debt story, not an AI conviction.

Then there is the study this series already leans on, which lands differently in an addiction post. METR's randomized trial (2025): experienced developers on repositories they knew deeply were 19% slower with AI — while estimating they were 20% faster. And METR's own 2026 update declared that result historical: its new measurement is unreliable. Partly because — read this twice — developers who feel dependent on AI refuse to enter a study that might take it away. The measurement problem is itself a symptom.

"My felt speed is the one variable the loop is designed to inflate."

And my own funnel? I had never counted it — a man who wrote three posts about measuring before believing had an unmeasured funnel in his own basement. So I counted, while publishing this post, on the public GitHub account. The readout says more than anything I could have narrated:

NEW PERSONAL REPOSITORIES PER YEAR · FORKS EXCLUDED Fourteen repos in six and a half months Public GitHub account "wifsimster", counted July 19, 2026 — 2026 is a half year 2 3 3 0 0 1 14 2020 2021 2022 2023 2024 2025 2026 three years, one repo — friction doing its sorting the agent era 3,665 public contributions in a year

Between February 2022 and May 2025: one single new personal repository in over three years. Three years of the friction doing its sorting. Since January 2026: fourteen new repositories in six and a half months — one every two weeks — and 3,665 public contributions over the last twelve months, the overwhelming majority since January, with spikes past a hundred a day. July to December 2025, on the same graph: near zero. You can literally see the month the brake let go.

The bottom of my funnel, counted honestly: of those fourteen repositories, about ten received commits within the last ten days — everything is alive at once, which is exactly what the loop predicts. One was consciously archived — it was called devteam; the irony comes free. One is "finished" in the most honest sense available: the birthday-invitation site served its event, and the event passed. The two January repos are still alive at six months — a small sample, but I'll take it. And "used by someone else": two repositories already pass the test — the-box, used by about ten people, and wawptn, by two. Two more are aiming for that bar: toko, in development, with me running the marketing solo, and copro-pilot, a long-haul build given the size of its scope. The count has its own trap, too: koe, quiet on the surface since June, actually lives inside the-box, copro-pilot and wawptn — a shared building block, not a dead project. solopilot, for its part, is just entering production, and its single user is a choice, not a failure: it is built for me only. For everything else, the user shares my name. Two real products out of fourteen, two candidates on their way: the funnel is harsh, but not empty. The global one above stays illustrative; mine now has numbers.

The structural reading is older than agents: code was never the bottleneck. Understanding, deciding, validating, distributing, maintaining — the bottom of the funnel — none of it accelerated. Velocity fills the top. The gap in the middle is exactly where the loop drops you: at the 90% mark, where the variable reward has already paid out and the remaining 10% is pure friction — the substance the addiction has taught me to avoid.

Second honest question: is it worth what it costs?

Diagnosis versus arithmetic, one last time — on myself.

The arithmetic (measurable). The subscription — mine is the Max 20× tier, $200 a month, the top of the consumer range. In API terms, an aggressive agentic month would cost well past it. The homelab's kilowatt-hours. That column is knowable to the dollar: $2,400 a year before electricity. (And symmetric skepticism applies: the viral claims that power users cost Anthropic $25,000 a month are unverifiable estimates of a private company's costs — an anti-vendor number deserves the same distrust as a vendor number.)

The diagnosis (real, unmeasurable). The evenings. The attention. The sleep edge. The 1970s house: its renovation, the hobby that used to produce finished rooms, has quietly stopped competing with a machine that pays out faster. None of that fits a spreadsheet, and all of it is cost.

The value, honestly, in three lines. First: learning. The method I forge on Sunday is the one my team runs on Monday — everything Solstice does at work was prototyped on my own time. That line is the most valuable, and the hardest to price. Second: own use — a homelab service I use daily is a finished product with one user, and it counts, as long as I don't dress it up as a business. The repo funnel doesn't even capture all of that value: agents also raised the quality of what already existed. Home Assistant now orchestrates far more automations and connected devices than before, and its dashboards were entirely reworked by agents. Upgrading the stock creates no new repository; it is still service delivered every day. Third: revenue — zero, like roughly half the indie universe, and that is only a problem if I claim otherwise.

And there is a value I did not see coming: the tool replaced the search engine. Every question that crosses my mind now goes through an agent, not through Google. I am not an isolated case — just ahead of the curve. StatCounter saw Google slip under 90% market share in late 2024, a first since 2015 — and StatCounter doesn't even count chatbot sessions. Gartner predicted in February 2024 that traditional search volume would drop 25% by 2026: the prediction missed, forget the number, the direction stands. The solid figure comes from Pew Research: across nearly 69,000 real searches analyzed in 2025, users clicked a result in 8% of cases when an AI summary appeared, versus 15% without. And to keep perspective: in mid-2026, Pew counts 49% of American adults who have ever used a chatbot — half the country never has — and 42% who use one to search for information. The reflex to search hasn't gone. The reflex to click has — and in my house, so has the reflex to open Google.

"The value is real. It is just not where the speed promises it is."

What the addiction actually costs

The tradition in this series is a section that lists what the post does not solve. This one gets the harder variant: the audit grid, applied to me. Behavioral science's most cited checklist (Griffiths, 2005) demands six components before "addiction" means anything: salience, mood modification, tolerance, withdrawal, conflict, relapse. All six are required — and the ICD-11's version adds one more condition: normally twelve months of functional impairment.

My honest scorecard: salience — in the shower, I am thinking about the agent still running — check. Mood modification — launching one relaxes me the way lighting the soldering iron used to — check. Tolerance — one agent used to be enough, I now run three in parallel; and where ten to twelve hours of development once made a heavy day, a day under fourteen hours of prompting has become an achievement — check, two proofs for one box. Conflict — some evenings, one stalled renovation — half a check, the honest half. Withdrawal, relapse — not really, not yet. Three and a half out of six, depending on the day's honesty.

"Officially, I am an enthusiast. Which is exactly what a gambler would say."

Two costs the grid doesn't capture. The idea compost: unbuilt ideas used to ferment, recombine, and come back better — the friction-era selection this post opened with. Now every idea gets built to 90% and nothing composts; I may be trading ten shipped mediocrities for the one good idea that needed a month in the dark. And the perception debt: I feel faster, and METR's developers felt 20% faster while being 19% slower. My feeling is precisely the variable the loop manipulates. That is not a reason to quit — it is the reason the next section exists. You cannot eyeball your own speedometer.

What to do Monday morning (and Sunday night)

Not a detox plan. I am not quitting — the machine answers 90% of what my brain produces, and I like it. These are the guardrails of someone who intends to keep playing:

  1. Count your funnel, monthly. Repos created, deployed, used by someone other than you, alive at six months. Four numbers, one line in a note. The count converts the feeling into a measurement — my own first readout is in this post, and it cost me one evening and two illusions.
  2. Set a shipping quota. No new repo while the last one is neither finished nor consciously archived. Starting is the reward the loop sells; make finishing the license to start.
  3. Apply the someone-else test. One user who is not you — a person, a colleague, a stranger — before a project earns the word "product". My homelab passes with one user: me. Name that honestly.
  4. Write "done" before the first prompt. One sentence: what exists when this is finished. An idea without a definition of done is a slot machine with better syntax.
  5. Protect one agent-free day. Not for virtue — for the compost. Ideas need unbuilt time to recombine. The soldering iron gave you that for free; now you have to schedule it.
  6. Total your real monthly cost — subscription, tokens, kilowatt-hours — and read it next to the funnel's bottom line once a quarter. Not to feel guilty: to decide with numbers what you currently decide with dopamine.

Where this leaves the series

Three days ago, the professional verdict: agents rebuilt how my team ships, and the numbers hold. Today, the personal bill: the same machine, pointed at one idea-rich brain with no product owner and no verification gate, produces a loop with no built-in exit. Same amplifier, both directions — that has been this series' claim since March. At work, the gate says no before a human reads the diff. At home, I am the gate, at 11 p.m., and the gate wants to see the next reward. The guardrails above are me building for my evenings what we built for the team: a stopping point that does not depend on willpower.

Read the rest of the series: The SDLC Is Dead — why the process was always waste. Coding Was Never the Job — what the engineer becomes when syntax disappears. AI Became a Team Reflex — what an AI-native team looks like on a Tuesday. DORA Just Confirmed What My Team Already Lived — the amplifier, measured. 100% Agent-Written — the professional verdict this post is the personal bill for. The Backlog Was Never the Job — the same shift, seen from the product role. The Tests Passed. Nothing Happened. — the defect that has no symptom, and what catches it instead of reading.

Sources