> If you’re using agents to program, what are you doing while they work?
If I am using agents I try to do something that is closely related to the task they are on. Otherwise I am just context switching once they are done and I want to review the work, which makes it difficult to focus on that task.
I also don't try to run too many agents at the same time as that is just madness. That's just herding cats at that point.
> As a side note, having Codex review Claude’s work (or vice-versa) throws up so many show-stopper issues (even with plan, revise, implement, review loops), I feel like you’d have to be nuts to just have a bunch of agents YOLOing it
Solely relying on agents is bad regardless. It certainly is the easy route and our brains are wired to take the easy/lazy approach. But even with how good models have gotten in the past year they still do make mistakes. In fact, they are now at a level where the hallucinations aren't obvious making it even more important to keep a close eye on the result.
If you do want to lean more heavily into agents doing most of the work, try to make sure they are following proper development practices. Something they don't actually do by default but using something like the superpowers skills makes a world of difference: https://github.com/obra/superpowers
Having them follow TDD helps a lot. I've even considered adding a QA agent/skill in here expanding things further to not just unit tests and some basic manual tests but also creating proper automated tests (following the test automation pyramid principles) to create an entire test suite.
Not to actually give more control to LLMs, but because it allows me to more easily spot where things go sideways. Since more tests, including e2e tests and UI tests where possible let me review more aspects of the work they do.
Having said that, I haven't created that skill yet. As reviewing the work of multiple agents is already exhausting as is. I am not in a position where I have to use AI or else in my company so instead I have dialed down my agentic usage by quite a lot to the point where I barely use agents anymore. To me using LLMs mostly as tools outside the process still is the sweet spot.