Where good ideas come from (for coding agents)

sunilpai.dev

3 pointsandric2 comments

I feel like I’m missing something between what is explained here and claims that agent driven coding massively speeds up output. Certain parts of this like Day 2s decomposition into steps and Day 3s context packet seem to imply that the author already knows the codebase. They know what steps need to be taken to fix the problem and they know what abstractions exist. That all makes sense to me and reflects what I’ve experienced. It helps get done what I already know. But I still have to understand and know the codebase which feels at odds with some of the massive upside some people are selling. This might be a little adjacent to the main point, but overall I’ve had similar good experiences with the points raised on day 2 + 3.

I touch upon this a little in the epilogue section https://sunilpai.dev/posts/seven-ways/#epilogue-ok-but-doesn... I was thinking what work there really is for "juniors" in this new future, and I suspect the answer then that their "job" is to develop the understanding of the codebase, approaching changes from the edges, and learning how to introduce new changes. (very similar to how people would onboard to a codebase pre-llms: by writing tests, doing smaller changes, poking about the code by adding console statements, etc). The good news is that the coding agent is a very good tool to do that as well. I plan on writing more about this soon.

Reply on news.ycombinator.com