In November 2025, everything changed. The AI revolution stepped into another gear with Opus 4.5. I remember the awe of spinning out projects with AI agents, the excitement, the empowerment. Yet almost a year later, I find myself deeply depressed.
I start my day and spin up my first agent. I plan the first issue and start cooking. Now I start another agent in a separate worktree. I do this about 4-5 times until I am bouncing between the many different agents. I feel a manic energy welling up. I’ve had to direct an agent about 6 times; it’s still failing. I need a skill to fix this behavior! I quickly open another agent to start developing a skill. I keep this up until, seemingly out of nowhere, it’s 5pm. I’m almost done, though! I push through till 7pm. At the end of the day, I’ve finished nothing. I’m 3 months behind in my project. The workflows that worked in the beginning have run into a brick wall. The pressure is mounting. It feels like one of those nightmares where a monster is stalking you. When you try to run, all you can do is spin.
Dopamine is foundational for human motivation and action. It is through dopamine that we are able to complete months-long projects at work. To work toward a career path over years and decades. To push through hard problems that seem beyond our capabilities. This same force also causes the social media doom scroll, online shopping, alcoholism, drug abuse, and gambling. The power of the dopamine cycle is indiscriminate with its application.
Our bodies want to be in homeostasis; in some ways Newton’s 3rd law applies just as well to human biology. Anna Lembke describes this phenomenon in her book Dopamine Nation as a seesaw. On one end is pleasure, and on the other is pain. The more we weight the pleasure side, the more our biology weights the pain side. The opposite is also true: as you push into challenging or painful experiences like cold exposure, strenuous exercise, or thinking extremely hard, your body responds with pleasure.
In gambling it’s not the reward that’s addicting, but the anticipation. The anticipation of the reward causes dopamine to rise.
So how does this apply to AI? Let’s look at what the product development cycle used to look like. Note that PAIN here is real work. Actually having to think hard. Having to make difficult decisions. Having to coordinate with people.
- Receive a PRD describing a project
- Project kickoff with the stakeholders
PAIN - Do an architecture meeting to plan how to solve the problem
PAIN - Break down the problem into smaller tasks
PAIN
Then loop through the tasks:
- Pick up a task
- Understand a task
PAIN - Read through the codebase and plan how to finish the task
PAIN - Spend a day or two coding the task by hand
PAIN - Put the code up for review
- Spend another day addressing feedback
PAIN - Merge the task. Launch to production
PLEASURE
Repeat until you finally finish the project, get a huge spike of dopamine, and feel a strong sense of accomplishment.
Notice that the ratio between pain and pleasure is weighted heavily to the side of pain. The work requires pushing my limits and thinking through difficult problems. My body responds with positive emotions.
Now contrast that with an example product workflow using agents.
- Receive a prototype or PRD from product
- Use an agent to plan the approach of the project
PLEASURE - Break the project down into smaller chunks
PLEASURE
Loop through the following:
- Pick up a task
- Enter plan mode. Answer the agent’s 5-10 questions
- Refine the plan
PAIN - Execute the plan
- Wait
ANTICIPATION - It went off course. Rewind, correct the prompt
ANTICIPATION - Wait
ANTICIPATION - Still wrong. Throw it away, start the prompt over
ANTICIPATION - Wait
ANTICIPATION - Holy shit, it one-shot the problem
PLEASURE - QA with agents
PLEASURE - Put up for review
- Point the agent at the comments to address
PLEASURE - Merge the task
PLEASURE
The agent loop strips out most of the pain of the old loop and replaces it with anticipation and pleasure, acting like a slot machine. The cycle is weighted heavily to the side of pleasure. My body responds with strong negative emotions.
Agents also make it seem possible to get on top of the unlimited tide of busywork each organization has. Difficult work still requires painful manual thinking. Why do the difficult work when I could pick up an easy task and get a shot of dopamine within an hour? Or break the task down into incredibly small bits? Or make a new agent skill? Or summarize the day using an agent? I pull the lever on the AI slot machine. Each time, my biology responds with pain. The mounting work overwhelms me, and the only way out is to PROMPT MORE. This is a vicious cycle. AI Depression.
The path out of addiction or depression is often the same. It requires abstinence or separation from the activity or substance. A focus on doing hard things and embracing the uncomfortable.
Taking a break from AI is unfortunately not something I can do. The adoption of AI is paired with an increase in work expectations to match. So that leaves me with doing hard things. But how can I do hard things when the agent has taken the hardship away?
I was entirely naive to AI when I began using it. It felt magical. The smarter models felt much smarter than I was. I began ceding my judgment. Second-guessing myself and letting AI take over the hard parts.
I often wonder how the technical staff of Anthropic move with such confidence and velocity. My hypothesis is that these are people who worked their way to the top of an industry. People who built discipline through building systems, working in large codebases and organizations. People who have confidence in their judgment. Maybe this team can use AI effectively because they are already disciplined thinkers.
Pushing onto the pain side of the day-to-day is only possible by focusing on the tasks that AI can’t do. By thinking hard about what to build and applying my taste.
My goal is to build this discipline in my day-to-day, to counteract the effects of AI. In practice, that means using paper and a pencil to think before I start an agent, limiting my AI usage throughout the day, keeping a portion of the day for analog work, and most importantly, starting a writing practice. This essay is rep one.
Being more disciplined is never a satisfying answer. Especially when the problem is rooted in our biology. I’m still not entirely sure that using AI 40 hours a week will ever be “safe.” Maybe the real answer is that the team at Anthropic is just as addicted as I am.