I use AI agents sparingly

In theory, AI agents do the legwork for me, but in practice, I find that I have to keep an eye on Claude — same goes for every other LLM. No matter how clearly and carefully, or in how much detail, I describe the tasks, it routinely finds a way to exercise good judgement. That’s not to say I can’t get top-notch work done without it, but I can’t trust it.

For example, if seeding your development database results in a foreign-key error, what makes sense?

  1. Correct the seed file.
  2. Correct the record in the development database.

Guess which one my buddy Claude chooses every time. (Sigh)

Nothing replaces human judgement, because ultimately the AI judgement is nothing of the kind, it’s simply an eager gollum behind a bloated React-driven TUI.



all tags