Comment by GoatInGrey

I'm unsure if this also qualifies as incompetence/embryonic understanding, though I've used LLMs for hundreds of hours on development tasks and have also found that sub-agents are not good at programming. They're more suitable for research tasks to provide informed context to the parent agent while isolating it from the token consumption which retrieving that context cost.

Zooming out, my findings on LLMs with programming is that they work well in specific patterns and quickly go to shit when completely unsupervised by a SME.

  * Prototyping

  * Scaffolding (i.e. write an endpoint that does X that I'll refine into a sustainable implementation myself)

  * Questions on the codebase that require open-ended searching 
 
  * Specific programming questions (i.e. "How do I make an HTTP call in ___ ?")

  * Idea generation ("List three approaches for how you'd ____" or "How would you refactor this package to separate concerns?")

The LLMs all fuck up on something in every task that they perform due to the intersection of operating on assumptions and working on large problem spaces. The amount of effort it takes to completely eliminate the presence of assumptions in the agent make the process slower than writing the code yourself. So people try to find the balance they're comfortable with.