A developer who maintains the open-source project LibreWeddingPlanner recently undertook a month-long experiment avoiding AI coding tools at work, documenting how the technology created unexpected problems despite initial efficiency gains.
According to the blog post, the developer initially embraced AI code generation features in VSCode and other tools, finding the speed appealing. Tasks that took days could be completed in hours. However, as usage escalated, he began delegating entire Jira tickets to AI agents, eventually connecting multiple agents to work simultaneously on different tasks across separate code repositories.
The expanding reliance on AI revealed several issues. The developer found himself unable to understand or explain the code being generated, creating anxiety around code reviews with more experienced colleagues. When tasked with reviewing AI-generated pull requests, he spent two days managing context-switching and revisions on work that AI completed in five minutes—turning apparent efficiency gains into actual time losses.
Quality control became problematic. According to the post, nearly every AI-generated pull request required modifications before merging. The developer noted that AI agents would fail to solve even half the tasks assigned, sometimes requiring him to mentor the system through repetitive corrections. He also observed what he characterized as degrading code quality over time, comparing the pattern to how Google allegedly reduces search result quality to increase ad impressions.
A critical realization emerged: over several months, the developer had written no code himself and hadn’t made a single commit directly. He described this as “laziness” and recognized he was losing fundamental technical skills.
The developer also questioned the economic model, noting that a single stalled AI agent spent $30 in tokens while providing nothing—costs he wasn’t directly responsible for. He suspects AI companies may intentionally degrade quality during uncertain financial times to increase token usage, similar to other tech industry practices.
The core problem, as described, was loss of control: reduced understanding of deployed code, decision-making abdicated to systems he didn’t fully trust, and cognitive exhaustion from managing multiple parallel workflows. The experience prompted his original decision to ban AI contributions from his open-source project months earlier.
Key facts
- The developer stopped writing code himself over several months while using AI agents
- Nearly every AI-generated pull request required human review and modifications before merging
- Context-switching between multiple AI agent tasks created exhaustion despite faster individual task completion
- The developer reports losing understanding of code deployed to production
- A single stalled AI agent cost $30 in tokens while producing no output