I went through it all

A developer who has extensively used AI coding tools including Codex, Cursor, Claude Code, Copilot, and GLM reports that despite roughly 5x productivity gains, human oversight remains essential. The engineer found that autonomous overnight agent runs produced unmanageable scope creep and spirals, and that AI bug triage cannot be trusted regardless of how many agents are chained together. The developer concludes that coding by hand is effectively dead, but AI still requires constant hand-holding and code review.

I went through it all. talk to ai. specs. epics. tickets. IDE's, cli's, tmux, codex, glm, cursor, claude code, copilot, whatever. The bottom line as of today, I still have to read the code. I prefer to iterate alongside with it. Triaging bugs is also tricky, you can't trust it's own triaging no matter to how many AI agents you pass it through . I left it a couple of times working at night autonomously and the result over morning were horrible it did what I asked it to do but the spirals, the scope creeps, are unmanagable, so I dont really do that anymore . No matter how many upfront guidance etc, it will stray away. you have to hold its hands - unless its a hobby project you really dont care how its built. Having said all it's its still ~X5 faster than doing it all by hand before AI era, and coding by hand is dead.