We just had a 2-day hackathon at my workplace. Our team built an impressive LLM based agent that does an initial triage of product issues (e.g. based on alerts). We had done some prep work before the main hackathon, but most of the coding was actually done during the two days. There were more than 100 commits across 7 developers. Very productive.

Cursor played a huge role in achieving this. So many parts of this system would have been a full fledged project of their own in earlier hackathons where we didn’t have this kind of GenAI assistance. For example, we decided to build a visualiser to show events as a timeline, so that the demo is more impactful. It then took us 1.5 hours to build this, with all features such as zoom, filtering, JSON upload and parsing, search, event interactions (tooltip, click for more info), etc. With Cursor, it was like each developer had their own “10x developer” assistant at their disposal. It took us this hackathon to realise that we could and should be using Cursor/Copilot much more. Having to work from scratch, and with unknown technology, forced our hand but in a good way.

The “10x developer” assistant is unreliable and gets confused though. When I asked Cursor to build the visualisation tool in a blank project, it was able to do it in one shot. However, to implement the same as part of the earlier UI (built on the first day), I had to direct and coax it incrementally - otherwise it did a really shoddy job. I attribute much of this to the nascent nature of this technology. Perhaps next year, it can be effective even when adding to existing codebase.

It was fun to see the vibe coding in practice. Like I mentioned in my earlier post on this subject, the elephant in the room was how it can cannibalise the need for software developers over time.