Back
One month without AI: a developer's account of dependency and regained control
SiTech AI Team3 წთ. საკითხავი

One month without AI: a developer's account of dependency and regained control

A developer has published a first-person essay about a month of deliberately working mostly without AI, describing creeping dependency, exhaustion and eroded skills, and how he regained control of the code he ships.

A developer who writes at blog.bustikiller.com published an essay on 25 September about a month of deliberately working mostly without AI, after several months in which AI agents had taken over almost all of his coding at work. It is a first-person account of creeping dependency, eroded skills and regained control.

Months earlier he had barred AI-generated contributions from LibreWeddingPlanner, the open source project he maintains on Codeberg, to take a position. At his workplace, where working with AI is routine, he kept using it.

The sweet start

It began with a friend's advice that once you start you cannot stop, and with the enhanced autocomplete in VS Code; code-generation models and agents followed. A long-time TDD practitioner, he first had the AI write the tests and then the code too, so the tests were no longer written independently of the implementation, which removes half the point of the method.

Losing control

Later he pasted whole Jira ticket descriptions into the agents and let them implement the work. On a codebase relatively new to him he lost track of which changes were really needed, and soon several agents ran at once, in separate git worktrees wired to Jira.

The accounting did not hold up. A task he could have done in 20 minutes took an agent 5 and then took him two days to review: reading code, tests and style, giving feedback, switching tasks, waiting for CI. Months passed before he realised he had not written a line of code himself, and that the AI solved fewer than half of the tasks he gave it.

The turning point

The decision came about a month ago in a code review, when a colleague noted that a test he had written did not test the scenario changed in the application code. The colleague was right. For a developer of more than 10 years of TDD, failing on something so basic was humiliating. Being pickier would not help, he concluded, because the AI eventually convinces you and makes you complacent again. He stopped, even at the risk of his job.

Taking over an AI-written draft pull request himself, he realised how little he understood of it, discarded the changes and started over. He went back to TDD and to pull requests touching five files instead of dozens.

What he says now

"Say no to drugs", he writes, half as a metaphor. The first step is to admit how dependent you are on a technology you do not control and that costs your employer a fortune. His fear of underperforming without the agents did not materialise: he says he still ships significant changes every day, gets pull requests merged, and can answer for an incident because he wrote the code. The essay ends with a short appeal: "Wake up. Be brave. Regain control."

SSiTech

SiTech — AI-powered web development

We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.