It seems using AI agents in a software engineering workflow s not going anywhere anytime soon.

Current opinions and questions

  • Where do we end up after not ‘coding’ for a while, do you lose the edge to do the thing yourself? And does that matter for the end result?
  • We probably still want to keep the cost of change low.
    • Software engineering skills and knowledge still matter.
  • TDD is more important than ever.
  • Architecture and design is more important than ever.
    • How to
  • How can we objectively measure if we ‘get more done’?
    • Or is it just a subjective experience?
  • I don’t find it enjoyable.
  • It’s probably good to be very specific on your tests and specifications when generating code.

Gathering research

  • Look up that one paper where experienced developers / repository owners think they were faster, but were actually 19% slower.
    • Is that because models weren’t as good some months ago? (2026-03-19)

Pitfalls because I’m human

And need/want to be in the loop

  • When there’s a lot of changes
    • I experience the tendency of not reviewing all lines but checking a pattern that was applied.
  • When the AI is busy doing things
    • I can’t really do anything else, the context switches still hurt.
  • It’s easy to be too generic and have it do a lot.
  • It seems I tend to stick to “agentic” engineering once I started with it, because
    • There’s a sunk cost fallacy.
    • Changing all the stuff generated by hand is very tedious.
      • I’m less intimately familiar with the code.

Lessons learned

  • It’s better to be very specific about a single test for example and describe exactly the scenario you want tested.
    • Review, then commit.
    • Further TDD practices.