i swear i saw a video this week of a tool that takes a natural language input and returns a pull request, but i can’t find it…
…anyhoo…it got me thinking:
the only thing more awesome than cursor (which i use every day) would be “agentic” cursor that can go all the way to a pull request.
think dosubot, but with all of the RAG-power of cursor, and actually returning a PR, not just some breadcrumbs for human devs.
feature request:
a “cursorbot” that is invited to my private GH repo, and attempts to solve issues by understanding the issue/query, creating/editing the code, and submitting a PR for human review.
not sure what the git integrations with vscode are, but potentially this idea doesn’t need to involve GH bots/actions, could just be a cursor mode that goes one step further to make new branches and commit code.
We are actively experimenting with different ways of integrating agents into the developer workflow. Would love to hear more about your experience with them!
i haven’t used aider enough to say where it might be better than cursor, but let me give you a sense of my vision:
right now i use cursor for a huge amount of my work: my personal blog, the software i’m developing for my company, data manipulation, text formatting, learning new packages and languages, etc.
as awesome as it is, all of these involve me being at my computer, with a decent internet connection, actively collaborating with cursor. it speeds me up, but i can’t be doing anything else with my time.
what i think would be even more powerful is to send cursor off on its own to try to solve the problem and submit a git commit, or even better, a full PR for me to review. there are a lot of simple use cases where i am fairly confident cursor could get it right on the first try, for example: “add a link to the pricing page to the login page”, “add a button that makes this element 100% width”, etc
my ideal workflow would be to be able to send cursor off on its own to take care of these small tasks, in a way that is integrated with git and compatible with our CI workflow. there are three potential outcomes:
cursor gets it right and i give the ok to push to master
cursor doesn’t quite get it right, and i need to edit the code, but it saved me a lot of time because i don’t need to get my brain oriented to the full context of the problem, i just need to fix its mistake
it completely gets it wrong, and i can either ask it to try a completely different approach, or in the worst case it is basically doing what dosubot does by helping find the relevant sections of code and suggesting one way to approach the problem. even if incorrect, this gets me to the outcome with less attention than if i started from scratch.
the main benefit is the ability for me to get a bunch of tasks off of my plate, and come back to QC them later when i have the attention to do so. in this sense, cursor could serve the functional roll of an intern or a junior dev.
And what size codebase are you thinking for these? Would you be okay with giving Cursor a general direction of where to make edits? (ex: make changes in the website/pages and website/components folders)
the codebase doesn’t need to be huge, and i think it would be totally fine to need to give cursor very clear instructions with @ references for where to look for files.
in my mind the MVP of this feature isn’t about making cursor any “smarter”, it’s about action. the inflection point is adding a mode where cursor goes all the way to a PR.
so it’s really just about the end action going one step further; instead of cursor telling me what to do, with the possibility of clicking the apply diff button to save myself some copy/paste, it is doing it in a new branch, committing the changes and making a PR.
fun workaround idea inspired by this (suspicious and smells like click bait) demo of gemini (from a google employee):
i’m going to get gemini to write a script that just uses the human ui of cursor for me…
(only half joking, been fooled too many times by these types of demos i have a hard time believing gemini could do something useful here, but i also sincerely hope someone proves me wrong because my arm hurts from clicking “apply diff” and “copy/paste”.)
I think directly working from the “project management” software is going to overtake the IDE’s. With sweep, I can be pretty productive making issues on Github from my phone
no, in my experience sweep is garbage and doesn’t do anything useful. the one cool thing they added was an open-source code-chunker that is pretty good for RAG with codeGen AI.