today i’m experimenting with just how large of tasks i can give Cursor Agent,...

today i’m experimenting with just how large of tasks i can give Cursor Agent, and…i haven’t found the upper bound.

it seems to be able to go off for 10-20 minutes at a time, do revisions on its own and find its own mistakes, and then come back with working code almost every time

it soon will get to the point where **i** am the weak point, where it takes so long for me to ask for more & merge results that it’s faster & higher quality to just have the AI do it all

that won’t happen all at once, it’ll be incrementally working higher up in the stack, until eventually..

It's a collaboration

i wish i could do “Constitutional Dev” with cursor. you sort of can with the system prompt. just a list of preferences, like

1. always use types
2. add unit tests for everything
3. tests must always pass
4. use plain SQL instead of ORMs
5. use default SQL values for stuff like IDs and timestamps..

it’s quite impressive, it’s even breaking a large task down into smaller chunks and testing each as we go

if i was a CTO, one big idea would be, how tf do i do this without sending anything to cursor/anthropic/openai??

like, im spilling my soul to cursor rn, and i dont see a reason why that would change. intelligence demands openness, thats always been a thing

the models are improving to the point where there’s a gaping hole ready to be filled:

who can standup a rack of GPUs in Frank’s old office so that none of our data leaves the building??

it gets me

at this point, the biggest cause of doom loops is Cursor’s bad UX, e.g. stupid me clicking buttons in the wrong order

damn, i’m the weak link? 🧐