I am trying to get a grasp of what we can do with agents. I am an avid user of agentic coding. But somehow I think I code faster by hand (by IDEs I mean). I've always had a sketch approach to coding. Where I kinda lay out the pieces of what needs to be done and then finish them in chunks. That way you keep building a mental picture of what needs doing. Lately I've noticed something - its quite fast to analyze a new code-base with agents for sure especially messy ones. But once I do get a hold of the codebase (even larger ones), I can change things by hand quite quickly. My bottleneck seems to be typing, but manipulating the mental map is a lot faster.
Does this match other folks experience or am I deluding myself ?
I will never be able to type out a large feature at even 1/100th of the speed of today's frontier LLMs -- even if I literally had a complete plan of every line of syntax I'd use.
now, the correctness of said feature is still up for debate, but we cannot argue that the thing came into existence faster than human hands could have wrought.
Yeah it’s way faster ime. Like a lot a lot, but there again I enjoy the slightly bigger picture (not too big though) so maybe I just cut syntax slow
typing speed was never my strong suit, I also was heavily into readable formatting which I often did by hand in a mixed team. so I'm not one of these stays-in-vim people that can pop out reams of code. in this way I appreciate the manual lift an LLM gives me.
I mean the question is 'faster at what'
You can vibe code something faster, but you have to know what your are doing to avoid making a big ball of mud. Anything non-trivial needs architecture that is 'better-than-current-frontier-model' capable.
Now tbh it's not clear at this point if the limit is the model capability or our ability to provide the right context to generate a good architecture, but for now it's one thing that you still need really smart engineers for when working on really hard problems.