Today I experimented with agentic development with Cursor after purchasing a Pro+ subscription costing $60 US, which pays for $60 US of tokens. The overall experience was relatively good, but checking the usage list made me think again whether Pro+ would suffice for any large project work.
Claude 4.5 Sonnet is really good, but burns $ at a rather rapid pace. For easy tasks, the custom AI agent Composer 1 by Cursor is rather sufficient.
Also, having long-winded sessions seems to push costs upwards due to ever increasing context.
Letting a second agent review what the first agent did and a third agent implement the requested fixes reduced costs to $0.50 for those steps. Compared to the roughly $20 I spent on the actual coding and debugging, that's a bargain. It was just about refactoring a small project to use a different library as dependency.
In the end, I didn't need to check the documentations of those libraries. But I needed to check the output of those agents very closely. They tend to drop functionality, if they feel it's difficult to maintain it!
#ai #dev