HN Explorer

Ask HN: What are the fundamentals of agentic coding?

1 points by guitarlimeo ↗ HN
This thought popped in my head today. Most of us probably know the fundamentals of programming by now - data structures, complexity (big O), type systems, functions, architecture etc.

But what are the fundamentals of this new era of coding? Have any had time to develop as it's changing so fast?

2 years ago prompt engineering was something people took seriously, but now the harnesses take care of it and you can again just prompt "its wrong please fix".

Maybe setting up feedback loops can be a fundamental? It seems to be the current fad at least.

Curious to hear your experiences.

7 comments

[ 2.7 ms ] story [ 12.2 ms ] thread
namanbnsl
I think one fundamental for getting agents to work well is giving them a number to hill-climb.

If you can create an eval for something in your project, the agent can keep trying different approaches until the score goes up.

I believe building good evals will matter more than ever.

verdverm
> just prompt "its wrong please fix".

Even though they can do more, the context engineering (in-repo markdown) is important. Think of it like BigO, where having good agents.md removes the need to re-deduce how code works, explains invariants, and the like.

My #1 fundamental right now is staying in the loop. Leaving the agents to their own devices is a recipe for bad things. Be skeptical of fads, these are typically the people who've stopped looking and are just finding new ways to manage slop without having to practice craft.

crog
- be curious - don't be lazy in your prompts - manage context - use sub agents - get 1 model from 1 lab coding, 1 model from another lab reviewing - ask claude and codex to call them in background (less tabs/visual clutter) - don't use too many skills - have fun, and improve because you're having too much fun and you want to iterate