Fully agreed, aligns perfectly with my experience hitting the 95% done wall on a solo contract project recently. I still do the majority of work using agentic tools but the multiplier effect feeling evaporated at a certain point as the accumulated tech debt, complexity and scope creep enabled by how “easy” features felt with Claude Code/Codex early on finally caught up to me.
I (probably) still would have used CC heavily with benefit of hindsight but with a view that every seemingly “trivial” feature CC adds in the greenfield stage to be radioactive tech debt as the pile grows over time. Until reaching the point where CC starts being unable to comprehend its own work and I have to plan out tedious large scale refactors to get the codebase into a state approaching long term maintainability.
It’s always tempting to start writing code before you really know what you’re going to build because it’s so satisfying and exciting to see an idea take shape. I know I’ve had more than one or two projects where I started writing before I understood the shape of the problem I was solving and ended up a few hours into the project with a useless pile of stupid. It seems like LLMs can lead you much further down that road because it just seems so magically productive.
I (probably) still would have used CC heavily with benefit of hindsight but with a view that every seemingly “trivial” feature CC adds in the greenfield stage to be radioactive tech debt as the pile grows over time. Until reaching the point where CC starts being unable to comprehend its own work and I have to plan out tedious large scale refactors to get the codebase into a state approaching long term maintainability.