TLDR 
Full circle as to how I used Cursor 2 years ago. I went down the deep end, now back at the good end. But this time I’m just using a clean model (Composer 2.5) trained for their harness and with much more speed and accuracy.
I have been a Cursor user for around 2 years now. I left for a few months around April/May 2026 for the Claude Code and Codex as the transition from Cursor 2 to 3 was abit buggy, and the amount of usage I got from Anthropic/OpenAI subscriptions was way more than Cursor gave (which makes sense). I am sorry for cheating on you HAHA!
Since Composer 2.5 paired with the new Cursor 3 layout has bought it back into the lead again, on simply a Pro or Pro+ sub, the Composer 2.5 usage is amazing value. It aligns now with also the way I have reverted to using agents. Still read the guides, docs, source code of packages I am going to be using, but have the agent do that instead and stuff the context window before going on.
I found when models like Opus and Sonnet and GPT 5 came around, it allowed features to be poured out very quickly, and due to my brains reward system getting super super hyped of seeing the new feature, it spurs on to do more, and then the core boring architecture never gets attention, just features poured in. Then I’d get bored, and start a new project HAHA! I fell into this pattern for a long time and now starting to come back to realisation I can still go slower with models like Composer 2.5 with more accuracy, but its still SOOO much faster than it was prior to Cursor existing. So, I am taking everything I have learned from this past 2 years and going to take a good effort at actually building something to stay and that people will use. Now, I am taking things slow, going back to basics as to how I would build an application before LLMs, but I no longer need to type a single line of code, which is outstanding.
It also has drawn me back to a very specific way of working with agents… which was just like I was when I first started using Cursor, minimal rules, getting it to reference existing code, using docker for accessing systems, etc. I felt I was getting abit lazy with LLMs and just saying “Add this”. But now if I actually take time to give it better prompts, example files, example data structures, I get great results with Composer 2.5.
I have 54% API credits usage on my account, and thought, damn, I am broke, I need to ONLY use those models (Opus and GPT) for SPECIAL cases, but I have not had to switch once which is great.
In regards to subagents, rules, they would very well in a large well architected codebase, with strict checks in place, but for new greenfield work, have fun 
as I think it would be chaos… haha!
The explorer subagents that break out automatically are great though
I am looking forward to whats coming with their next models and believe they have the capability to fully take on OpenAI and Anthropic in terms of coding ability. It just makes sense given the sheer compute power they now have, the talent and the data that ppl shared with them, specific to this harness.
