I’ve been experimenting with this idea that I can implement features faster and more accurately than the big model + agent swarm paradigm. I also have been trying this because of what the post brings up when it comes to AI burnout.
For me, writing the code (literally typing), shifting files around, renaming things is slow. So I’ve been using small models + much stronger grasp on the reins. Similar to what this post mentions. It’s been great, the decisions come from me. I tell it to make a domain class with these fields and these invariants. It does it in a split second, I look it over, tweak it and move to the next part of implementation.
When I experimented with swarms it would spend an hour just having agents adversarially review to decide some pretty trivial details. This way, when someone asks me a question during review, I can answer it. When I need to dive back in, I know what to look for and where to look for it. I have way more connection to my work, than a few months prior.
Early copilot was exactly what I wanted for a long time. A smarter autocomplete. Then they started taking it too far with just filling in the entire file.
It's the difference between constantly reviewing someone else's code and getting to write the code yourself.
There's both deep satisfaction and deep value in having an understanding of the past, present, and notional future of the system (especially at scale). If I use agentic tools to do the actual coding, I never feel like I have as deep an understanding of the implementation, it's inherent assumptions or lessons learned. Over time, my understanding of the system atrophies badly enough that it becomes difficult to steer the LLM, instead of the other way around.
This is what makes those "unicorn"/great managers who both manage well AND maintain a clear understanding of the system over time so amazing.
cautiouscat · · focus · HN ↗
For me, writing the code (literally typing), shifting files around, renaming things is slow. So I’ve been using small models + much stronger grasp on the reins. Similar to what this post mentions. It’s been great, the decisions come from me. I tell it to make a domain class with these fields and these invariants. It does it in a split second, I look it over, tweak it and move to the next part of implementation.
When I experimented with swarms it would spend an hour just having agents adversarially review to decide some pretty trivial details. This way, when someone asks me a question during review, I can answer it. When I need to dive back in, I know what to look for and where to look for it. I have way more connection to my work, than a few months prior.
meander_water · · focus · HN ↗
And that's not a criticism, that's what I've come back to myself.
dawnerd · · focus · HN ↗
classified · · focus · HN ↗
Did it ever occur to you that slowness is good because it gives you time to think?
aaaronic · · focus · HN ↗
There's both deep satisfaction and deep value in having an understanding of the past, present, and notional future of the system (especially at scale). If I use agentic tools to do the actual coding, I never feel like I have as deep an understanding of the implementation, it's inherent assumptions or lessons learned. Over time, my understanding of the system atrophies badly enough that it becomes difficult to steer the LLM, instead of the other way around.
This is what makes those "unicorn"/great managers who both manage well AND maintain a clear understanding of the system over time so amazing.