Ai-architecture-policy-and-costs-reduction-roadmap.pdf (20.8 KB)
I’ve been successfully skyrocketing (I mean it) the lines of codes I produce monthly on magento 2 and custom react/ts projects and I’ve been discussing with my employer where this may take us.
Basically the 5.2-codex model I’ve been using costs us (for me only, single developer) up to 1.000 $ per month. we pay for development success and great performance and any penny of them is definitely worth it in my opinion: we already see customers sharing and driving the idea.
that said, we envision a future where cost awareness could be key, and for me Cursor can be a key toward success.
that said, before embarking in working with local models (see attachment if interested on why) I’d like to know how cursor team envisions this.
I’m asking this now because cursor 2.4 bring in an exiting idea: subagents.
I’d like to know what cursor team or other developers think about this: think for example at the company I’ve been working for: 15 devs x 1.000 bucks dev/month (and just scrapping the surface of cursor potentiality…)… this is something to take into consideration. I see this as a paradigm change.
having possibility to have local models that subagents can levarage for basic tasks (recurring cmds, retries/iterations, search, indexing, search and replace, application standard cmds) could be an interesting balance.
cost reduction is not the issue here: awareness for me is the key going forward. I see great potentiality, I’m a programmer by trade but an experienced linux sysadmin hence using an “old mining rig” for an ia local infrastructure (pure testing I mean) could be cheap but effective: I’ve been striving to understand NOT really how to install and make the local ollama (citing it just as an example) operational, but rather this: is there a field that could really make a local model shine? reindex? task-specific subagents? and: how to configure cursor to orchestrate it correctly or the best it can?
and: does cursor team see this as a possible part of its future infrastructure environment (means: any clear roadmap for support going forward)? I want to be extra careful in investing time on this if the cursor dev team honestly tells me “we see where it’s going, you’re not going to get anything out of this honestly” or “we’re gonna drop support for local models in the future”.
I’d love to know if cursor team has plans for support and if someone has got genuine ideas about a starting point here.
to summarize the attached analysis: “The goal is not to replace frontier models, but to use them intentionally, while offloading high‑volume, repeatable work to local infrastructure.” and “Most Tokens Are Not Deep Reasoning”.
not sure about the last part: but the 5.2 model suggests this: very interesting to me.
thanks