A case for sequential work

As it has been lately, I’ve been messing with writing software with AI. Experimenting with harnesses, agents, remote machines - all of it.

As with most things, the human desire to experiment with “more” is strong. After all, more is more and more is better, right?

Well .. not really, no. And that’s no surprise. Anyone that has tried to push a limit knows, there are diminishing returns. It doesn’t mean it’s bad or you shouldn’t do it. It just is.

With regard to writing code - in practice, I’m seeing very little good at this stage from trying to do work in parallel. Having a good plan, well broken down work - all of that is good and valuable. My assumption being that, by having all these nice things we can do more, faster. In the end though, all the communication and synchronisation overhead makes for more work.

Parallel reads, summarisation - those are a different story. For that - sure. Writing code though, I’m not convinced you end up with better results. You get the same at best. But the token burn is much larger due to all the additional costs. Maybe it’s faster, maybe. I’ll be honest, I’ve not bothered measuring, but I think it’s going to change my opinion on it. Also - watching 100s of subagents spazzing out on in a recessive fashion feels bad. One agent chugging along for hours on a well planned work - kinda chill vibes.