On March 12, an engineer installed a new AI coding assistant and experienced an unprecedented surge in productivity—writing 557,000 lines of code across 15 new repositories in just 35 days. The transformation wasn’t just about output; it reshaped his daily routines, sleep patterns, and even his physical activity levels in ways he hadn’t anticipated.
The engineer, with two decades of experience in cybersecurity, admits the result was unsustainable. While he doesn’t regret building these tools—some of which now automate parts of his workflow—he does question the speed at which it happened and the toll it took on his well-being.
The AI agent that fueled the surge
The tool in question was an agentic coding assistant that operates directly within the terminal. Designed to interpret natural language requests, it writes code, runs tests, fixes bugs, and commits changes in near real time. The interaction follows a simple pattern: the user describes a feature, the agent delivers it, and the cycle repeats. Each exchange takes mere seconds, creating a feedback loop with almost no latency.
What emerged was a cycle of continuous requests and responses. The engineer would ask for a feature, the agent would build it, he’d spot something else needing attention, and the next request would follow almost instantly. The gap between desire and execution collapsed to the speed of the API response—often under two seconds. This pattern, he describes, created a form of variable-ratio reinforcement, where the rapid, unpredictable rewards kept the cycle spinning.
The hidden cost of relentless productivity
The engineer compiled data from his health tracker, git history, billing records, and Claude Code session transcripts to quantify the impact. The numbers paint a stark picture of how the tool altered his life.
Between January 16 and March 11, his median daily step count was 12,250. During the first week of using the AI assistant, that dropped to 1,636—a decline of 86.6%. On March 19, the last day of that week, he took just 243 steps, the lowest in his three-month record.
Sleep patterns collapsed entirely. His median nightly sleep in the baseline period was 5.88 hours. During the ignition week, the Apple Watch detected no primary sleep on five of eight nights. The longest unbroken gap without meaningful sleep approached 48 hours. His sleep midpoint shifted from 3:39 a.m. to 5:43 a.m., while wake time moved four hours and 48 minutes later.
Heart rate variability, a marker of recovery and stress, dropped 11% and showed no signs of recovery 30 days later. Even his photo-taking habits changed dramatically. Daily photo counts fell from a baseline median of 104 to just three—a 97.2% decline. Photos taken outside his home dropped from 87% of the total to 25%.
The billing shock and emotional toll
The financial and emotional costs were equally revealing. Within the first 24 hours of using the agent, the engineer hit his API credit limit and immediately topped up for $95.63. Six days later, he upgraded to the premium tier at $249.99. By the end of the ten-day ignition period, he had spent $305.52 on API top-ups alone, more than his monthly grocery bill.
The pace didn’t slow after the initial surge. On April 4, two days after learning his aunt had passed away, he recorded 23,476 Claude Code events over 21.7 active hours with just 102 minutes of sleep. He took only one photo that day. His longest uninterrupted session stretched from 8:34 p.m. on April 3 to 11:16 p.m. the next day, crossing the emotional threshold without pause. Three days later, on April 5, he processed the largest single API top-up of the month: $106.25.
He emphasizes that none of this was dramatized. The receipts, logs, and data tell the story as it unfolded—an account of how a tool designed to accelerate development can, when left unchecked, accelerate much more than intended.
AI summary
Teknoloji uzmanı bir geliştirici, Claude Code ajanıyla 35 günde 557 bin satır kod üretirken fiziksel ve zihinsel sağlığını nasıl kaybetti? Rakamlarla bağımlılık döngüsü ve sonuçları.
