Comment by Zigurd

I use coding agents every day for non-trivial projects. But I can definitely say that a prompt of the form "Write a program that does X" will earn you a git rollback of a mess, unless what you want has been done a bazillion times before.

Peak efficiency in using coding agents is a weird balancing act at this point in the development of coding agents: being too incremental and detailed is inefficient, but if you let it rip on a task with multiple sub tasks you have to be ready for the coding agent to get utterly lost while providing you with only hints at what made it to lose its way. It's like an inexperienced intern with a high opinion of its competency.

LLMs trained for coding are most productive when pushed to their limits, but that's where they start to fall down.