As I’m slowly gaining experience working with LLM coding agents I’m noticing some patterns in my work that are problematic and if not kept in check will probably become serious issues.

  • frequent pausing and switching context
    • between prompts, working on multiple topics
  • scope and feature creep
    • the LLM is incredibly thorough and easily derails the work
      • subtle bugs it found that were unknown previously
      • related issues
      • edge cases that would almost never happen
  • lack of attention, neglect
    • not reviewing and pushing back on overengineered results
  • lack of comprehension and learning
    • letting code slip through without understanding it

Previously, skill devlopment and learning were bundled with the process of creating a software product. You write, review, debug code and you learn. Now learning and creating are unbundled.

Tech Veteran - https://www.youtube.com/watch?v=pQPUWJK229c