Cybernetic enhancement, light speed, and the new purity of dopamine

On a weekend alone, four projects, two hundred and twenty two commits, and the strange realization that I had found a new, very pure source of dopamine.

My fiancee left for a work trip a few weekends ago. Three days, maybe four. I had the house to myself, no one to check in with, no plans on the calendar. By the time she walked back through the door, I had spun up three brand new projects (daybook, localPress, and a Humble Bundle key scraper), shipped v2 of Doorman, and put a real dent in the new TUI for Coco. Strut and my portfolio site got incidental polish along the way. One of those days I crossed 222 commits on GitHub. I do not say that to brag. I say it because when she got home and asked how my weekend was, I felt like I was coming out of a time capsule. Overstimulated. Jittery. Like I had been mainlining espresso for 72 hours. I had not meditated in two days. My brain was still in delegation mode, queueing up the next thing before it had finished registering she was in the room.

That was the moment I had to be honest with myself about what agentic coding is doing to my attention, my routines, and the way I get my fix.

The shift

I have been using AI coding tools for a long time. I was poking at OpenHands back when GPT 3 had just dropped and Llama 2 was the local model worth tinkering with. Two plus years of watching this stack evolve from “autocomplete with vibes” to “I can hand you a chunk of work, walk away, and come back to a pull request.” That last bit is the part that has rewired me.

Somewhere in the last year, the tooling crossed a threshold. The tasks I can dispatch have gotten longer and more autonomous. I can kick off an agent on my laptop, leave the house, and check in from my phone an hour later. I can queue up an overnight task before bed and wake up to a branch ready for review. The unit of work I delegate is no longer “write this function.” It is closer to “go figure this out and report back.”

What I noticed that weekend, and what I keep noticing, is that my planning brain never turns off anymore. There is always a next thing to queue up. A bug to dispatch. A refactor to send out. Going to sleep used to mean closing the laptop. Now it means picking which long running task to kick off before bed so I can wake up to fresh output. I am not exaggerating when I say I lie in bed thinking about what to point the agents at next.

Why my brain, in particular

I should name something up front. I have ADHD. My brain is wired toward novelty and exploration in ways that are pretty well documented at this point. Research on ADHD and dopamine has consistently found that ADHD brains run lower baseline dopamine and respond strongly to novel, unpredictable rewards. A 2018 paper in the journal Brain models this directly, looking at how novelty processing and reward learning are altered in ADHD and how stimulant medication shifts the picture. There is even a name for the gene variant most associated with this trait, DRD4, sometimes called the “wanderlust gene” in evolutionary psychology papers because of its association with novelty seeking, impulsivity, and a strong preference for exploration.

What that means in practice, for me, is that I have always been the person who finishes the prototype and immediately wants to start the next thing. The interest curve drops fast once a problem is solved in my head, even if the actual implementation still has weeks of work left. Agentic tooling did not create that pattern. What it did was remove the cooldown. The slow part of building used to act as a natural governor on how often I could chase the next novel thing. Now there is not one.

A new purity of dopamine

My brain loves problem solving, building, shipping. It loves the moment when an idea becomes a working thing. Before agentic tooling, there was a natural pacing to that loop: think, type, debug, ship. The dopamine hit at the end was earned by hours of slow grind in the middle.

Agentic tooling collapses the middle. I think, I delegate, I review, I ship. The dopamine hit lands almost as fast as the planning thought. It is the same chemical reward I used to get from finishing a feature, except now I get it at the pace I used to get from finishing a function. Sometimes faster.

I am not going to call this addiction, because I do not think that is what is happening. But I am going to be honest and say it is the cleanest, purest hit of the “I made a thing” feeling I have ever had access to. And I have access to it 24/7, from my phone, while I am brushing my teeth. That is a different relationship with my computer than I had two years ago, and I am still figuring out what to do with it.

The framing for this is not new. Anna Lembke, the Stanford psychiatrist who wrote Dopamine Nation, has been making the case for years that we are living in a time of unprecedented access to high reward, high dopamine stimuli, and that the smartphone in particular acts like a hypodermic needle for that kind of stimulus. Her work mostly focuses on social media, gambling, food, and the obvious behavioral targets. What I would add, from inside this particular workflow, is that agentic coding for someone whose dopamine system already runs hot belongs on that list too. The hit is cleaner, the loop is shorter, and the output is genuinely useful, which makes the whole thing harder to recognize as the same shape of compulsion. Useful work is a much better wrapper for a dopamine loop than doomscrolling, and I mean that both as a defense and as a warning.

Cybernetic enhancement

The framing I keep coming back to in conversation is that this is a cybernetic enhancement. Not a tool I pick up and put down, but an extension of how I think and operate. My capacity for parallel work has gone up because I have offloaded the part of building that used to bottleneck me. The actual typing. The boilerplate. The “I know how to do this, I just have to spend an hour doing it” part.

The upside is real and I am not going to pretend otherwise. I shipped a stupid amount of useful work that weekend. New tools that other people can use. Real features in projects that needed them. None of that is fake or hollow, even if the velocity is suspicious.

The downside is that my body and my attention are still made of the same stuff they were before the enhancement. I can dispatch faster than I can rest. I can plan faster than I can be present. The wiring has been upgraded, but the host is still a human who needs to eat, sleep, see his partner, and occasionally stop thinking about software.

The wisdom piece

Here is the part that I think gets lost in the outside view of all this. From the outside, agentic coding looks like “anyone can write code now, just tell it what you want.” That is not wrong, exactly, but it is incomplete in a way that matters.

The metaphor I keep reaching for is light speed travel. If you are traveling at normal speed and you are two degrees off course, you notice and correct. If you are traveling at light speed and you are two degrees off course, you end up in a different solar system. The faster you go, the more expensive a small misalignment becomes. And the harder it is to notice you are off course at all, because everything is moving so fast that “this is fine” looks identical to “this is catastrophically wrong, just not yet.”

Recent example from my own work. I was building out a Trigger.dev automation platform for an organization. The idea was to wire up their CRM, Notion, WhatsApp, and personnel coordination into a single organizational brain, with an AI layer doing aggregation and summarization. In the brainstorming phase, one of the early bullet points read, more or less, “LLM has KV store memory.” One line. Single bullet in a longer list. We moved past it.

That assumption got baked into roughly fifteen workflow specs. All of them implicitly relied on having persistent memory the LLM could read and write to between runs. When we got to deployment, surprise: there was no KV store available in the constrained infrastructure we were targeting. The line item that nobody had pressure tested in the brainstorm cascaded into a wall of broken first deploys. We ended up reworking the approach to use a last-sync timestamp block written programmatically into the Notion pages, which got us functionally equivalent state persistence, but only after eating the cost of every workflow that had been built on the original assumption.

The lesson is not “AI made a mistake.” The lesson is that a single ambiguous line in a spec, at light speed, becomes fifteen broken deploys. The wisdom that catches “wait, what do we actually mean by KV store, do we have one of those in this environment, and if not what is the fallback” is the wisdom that comes from having shipped enough things to know which bullet points are load bearing and which are decoration. That wisdom does not come for free. It comes from years of doing it the slow way, watching things break, and remembering why.

So when the outside view says “anyone can build software now,” what I want to add, gently, is: anyone can dispatch a build. Knowing what to build, and knowing which load bearing assumptions need to be tested before you press go, is still a craft. It is a craft that is easier than ever to skip, because the tooling will happily run with whatever you give it. And the cost of skipping it scales with how fast you are moving.

There is a study I keep thinking about. In July 2025, a nonprofit called METR published a randomized controlled trial on experienced open source developers using Cursor with Claude 3.5 and 3.7 Sonnet. Before the study, the developers predicted that AI tooling would make them about 24% faster. After completing the tasks, they self reported feeling about 20% faster. The actual measured result was that they were 19% slower. They felt fast and they were slow, and even after being shown the data, many of them still felt fast. Then in early 2026, METR tried to run a follow up study with a larger pool of developers and the newer tooling, and ran into a problem they did not expect. Too many developers refused to participate because they did not want to work without AI for the duration of the experiment. The data set got too biased to publish.

Two things in that. First, the feeling of speed is not the same as speed, and the gap between them is wide enough that even motivated participants in a controlled trial cannot reliably tell which side of it they are on. Second, the dependency that developed between mid 2025 and early 2026 was strong enough to compromise a peer level scientific instrument. That second one is what stops me when I think about it. The feeling of having this stack available is now load bearing in how a lot of us experience our work, regardless of what the stopwatch says.

The skill atrophy question (and the case against it)

There is a related strand of research I want to flag, because it shows up in the wisdom argument in a way I find uncomfortable. A 2025 study in the journal Societies, surveying 666 participants across age groups, found a significant negative correlation between frequent AI tool usage and critical thinking skills, mediated by what the researchers call cognitive offloading. The effect was strongest in younger users. Higher education levels seemed to insulate against it.

I do not want to overread a single study. The author flags it as correlation not causation, and there is a real possibility that people with weaker critical thinking skills are simply more likely to reach for AI tools to begin with. But the mechanism is plausible and it lines up with how skill atrophy works in every other domain. If the muscle that catches a load bearing assumption in a brainstorm is exactly the muscle I am increasingly outsourcing, I should probably make sure I am still doing reps with it.

Here is where I want to push back on the strong version of that argument, though, because I do not actually think dispatching work to agents has been making me dumber. If anything it has been the opposite, and I want to be specific about why.

When I am actually paying attention during a dispatch, the agent is constantly forcing me to think at one level of abstraction higher than I usually would. I have to specify what I want clearly enough that another agent can act on it. I have to anticipate the load bearing assumptions, because if I miss one, I am going to feel it in the output. I have to read code I did not write and decide whether it is correct, which is a deeply different skill from writing it myself. And when the agent does something I would not have done, I either learn why their approach is better, or I learn why mine was, and either way I have done a small piece of comparative review I would not have done solo.

I have learned more about systems design, infrastructure patterns, and language idioms outside my main stack in the last year than I did in the three before it, and almost all of that has come from being engaged in agentic loops rather than checked out of them. The risk in the atrophy research is real, but it is contingent on the user being a passenger. If you are driving, the same workflow is one of the densest self teaching environments I have ever worked in. The variable is attention, not the tool.

What I am doing to stay sane

The honest answer to “how do you keep this from eating you” is that I do not have it fully figured out. But here is what I have noticed actually works for me.

Daily meditation, somewhere between 10 and 30 minutes, sometimes longer. My partner and I do this together when we can, which makes it harder to skip and turns it into a shared anchor instead of another solo discipline.

Ending the day with something physical. Gym, yard work, anything where the feedback loop is muscles and not pixels. The body has a different sense of pace than the brain on agentic coffee, and listening to the body forces the brain to slow down.

The workshop above my garage. This is the one I have come to lean on the most. I have been getting deeper into woodworking and making physical things, and the pace there is just set. Wood does not care that I can dispatch fifteen parallel agents. A miter cut takes as long as it takes. Glue dries on its own schedule. There is no overnight task I can queue up that delivers a finished cabinet in the morning. That forced pacing is a feature, not a bug, and it is doing more for my nervous system than I expected.

(The irony is not lost on me that I am currently thinking about building or buying a CNC machine for the workshop, so I can do programmatically precise cuts on my own designs. I am aware I am trying to import the agentic workflow into woodworking. I am going to do it anyway. We can debate whether that counts as cheating in a future post.)

The last thing, which is the hardest, is being mindful about how much I let my brain stay in build mode after the work day is technically done. The side projects and open source work are real things I love. They are also the easiest place for “just queue up one more thing” to eat an evening. Catching myself doing that, and choosing to put the laptop down, is a practice. I am not always good at it. I am better than I was a year ago.

The open question

What I am genuinely curious about is how other people, especially other early adopters and especially other folks with brains that run hot, are navigating this. Not the productivity side. The other side. The “I noticed I am thinking about my next agent dispatch while brushing my teeth” side.

How is the satisfaction of shipping holding up for you when shipping is this cheap? Are you finding that the value of any single feature feels lighter than it used to, because there is always more queued up behind it? How are the people in your life experiencing this version of you? How are non engineers around you evaluating the work, given the outside view that “AI did it”?

I do not have clean answers. I have a meditation practice, a workshop, a partner who notices when I have been in my head too long, and a working theory that the wisdom of doing things the slow way is more valuable now, not less. The rest I am still working out. If any of this resonates, I would love to hear how you are working it out too.

Sources and further reading

• Becker, Rush, Barnes, Rein. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, 2025. The randomized controlled trial showing experienced developers were 19% slower with AI tooling while feeling 20% faster.
• METR. We are Changing our Developer Productivity Experiment Design, February 2026. The follow up note explaining why a planned 2026 study could not produce a clean signal, since too many developers declined to work without AI tooling.
• Gerlich, M. AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 2025. Survey of 666 participants finding a negative correlation between frequent AI tool usage and critical thinking, mediated by cognitive offloading.
• Sethi, A. et al. A neurocomputational account of reward and novelty processing and effects of psychostimulants in attention deficit hyperactivity disorder. Brain, 2018. On novelty seeking, dopamine, and reward learning in ADHD.
• Lembke, A. Dopamine Nation: Finding Balance in the Age of Indulgence. Dutton, 2021. The book that argues for self binding and intentional friction as tools for managing high dopamine environments.