Brad Zacher

15.5B tokens. 266 pull requests. 3,878,930 lines added. 5,278,449 lines removed. First month at Cursor and it was big, to say the least. It's the first time I've really leant into agents and needless to say you really can get a lot done. I landed a LOT of build infra clean ups.

Used an agent for profiling for the first time today. It was pretty cool to let it just run some tools and try to find some potential improvements. It came up with some vague suggestions I was able to grow and iterate on for some decent wins!

Pro: I got some (okay) recommendations for work, quickly. The good leads mostly just confirmed suspicions I had, but it felt good to get a 2nd opinion. Con: Many of the recommendations were obviously flat-out wrong -- if I wasn't as experienced I might have wasted a lot of time!

Used an agent for profiling for the first time today. It was pretty cool to let it just run some tools and try to find some potential improvements. It came up with some vague suggestions I was able to grow and iterate on for some decent wins!

If I ever see someone's comment and it says "you're absolutely right" I now automatically assume that they are either an LLM bot or that they're copy pasting from one. It's such a shame that we have to deal with this slop contribution in our opensource projects.

This isn't just limited to PR either - we've seen the same behaviour with issues. I recently saw one user file a few issues that were clear and obvious output from asking the agent to plan out a piece of work. They filed an issue with well over 1000 words! And the responses...

If I ever see someone's comment and it says "you're absolutely right" I now automatically assume that they are either an LLM bot or that they're copy pasting from one. It's such a shame that we have to deal with this slop contribution in our opensource projects.

One major issue I've found with package.json imports over tsconfig paths is usage across "boundaries". A package.json import spec cannot include a file outside the current dir -- eg it can't start with `../`! They also don't "bubble" through nested package.jsons!

If you have a workspace you can't define package.json imports in the root and have them work in all your child packages. Nor can you have the child packages access paths from the sibling packages. This, OFC, makes sense to avoid publishing broken npm packages. But sucks for non-publishing projects

Been seeing an uptick in what are clearly vibe-coded PRs. On the surface this seems great -- more contributors! But in the reality it's often AI slop that the person hasn't reviewed and instead the burden of review is on us, the maintainers -- which is a horrible trade.

I asked cursor's AI agent to try to create a small TUI form. It went HORRIBLY. After a while it declared the task as done despite numerous compiler errors. When I told it that it wasn't done. This was the agent's response:

I've been doing a bunch of work on a meta-linter of sorts and so I need to ingest results from many different linters. A really painful problem I've discovered recently is that nobody is consistent about how to define a "location" in text. It's a surprisingly complex space to try and reconcile!

The other report style is a line + column pair. Immediately we run into the same issues: - are the two numbers 0-indexed or 1-indexed? Thankfully the definition of a "line" is consistent from what I've seen. But a column is just another ill-defined index with no consistent definition 😭

I've been doing a bunch of work on a meta-linter of sorts and so I need to ingest results from many different linters. A really painful problem I've discovered recently is that nobody is consistent about how to define a "location" in text. It's a surprisingly complex space to try and reconcile!

Some linters report as a position -- a number representing an index from the start of the file. Seems simple enough but we immediately run into our first problem: is it a 0-index or a 1-index..? Few linters clearly define this yet it can completely break your reports and fixes!