Same for me. This also gives be just enough „hands on“ so I still understand what’s going on. It’s also the most fun.
I do like making software. I also like thinking about UX.
None of thise works (for me):
- research and plan a huge feature or entire service and let a fleet of agents implement it (agents get lost in slop)
- hold agents by the hand and micromanage them (too slow, inefficient)
- drown agents in rules and guardrails (effectively the previous point)
- let agents roam freely (vibe coding, good for throwaway software - but doesn't work for "regular" software)
What does work:
- a set of reasonable guardrails, rules, linting, a good testing strategy.
- decently sized tasks with research + planning upfront
- work your way through a problem task by task, layer by layer - doing work in parallel if possible
It's been a process but I'm super happy with how these latest models perform with this approach