Damus
jb55 · 1d
not from my experience. It may write better code locally but it tends to have amnesia globally, leading to lots of duplicate functions. Refactor prompts have been productive for this
nostrich profile picture
💯. For anything with nontrivial implications (eg in a lightning node!) they still suck compared to humans. Self-contained or straightforward tasks they’re absolutely great, but the limited context windows and where there are multiple potential directions they’re not great.
1
Vitor Pamplona · 22h
That just means that your Skills.MD, Agents.MD, etc file is not good enough. What I noticed is that in large codebases, AI needs a "summary" of what exists in the repo because it cannot send the entire codebase in the context window and searching for things all the time is extremely slow/costly.
jb55 · 22h
I think this is why agentic search is so good, if it doesn’t have enough context it just greps for what it needs