Developer admits 60% of his codebase is now unreadable after full LLM delegation

A software developer has shared how his use of large language models evolved from simple code queries to fully automated multi-agent orchestrations that write and merge code without his direct review. He initially drew a firm line between LLM-assisted development, where each increment is manually reviewed, and 'vibecoding', where output goes unchecked, vowing to stick to the former after three failed vibecoding attempts. However, as he progressively delegated more tasks to specialized agents and parallel workflows, his original resolution quietly collapsed. He now estimates that 60% of his codebase has become unreadable to a human, with that share growing after every orchestration cycle. Despite acknowledging the dependency risk — including the difficulty of reclaiming his own code if LLM access were cut off — he admits that abandoning these tools is not something he is considering.