I am, oddly, able to get really quite a lot of mileage out of $20/mo of OpenAI plan, and I have never encountered a usage limit. I have gotten warnings that I was close a couple times.
I wonder what I’m doing differently.
I did spend quite a bit of time, mostly manually, improving development processes such that the agent could effectively check its work. This made a difference between the agent mostly not working and mostly working. Maybe if I had instead spent gobs of money it would have worked output tooling improvements?
I wonder if you're like me? I tried out the MCPs and sub agents and rules and bells and whistles and always just came back to a plain Codex / Claude Code / Cursor Agent terminal window, where I say what I want, @ a few files, let it rip, check the diff, ask for some adjustments, then commit and start the process over after clearing context.
Haven't found a process that beats this yet and I burn very few tokens this way.
I don’t really write code with it at all, and that’s why I burn so many tokens.
I like writing code, I’m good at writing code. What I hate doing is dredging through logs, filtering out test scenarios and putting together disparate information from knowledge silos - so I have the AI doing that. It’s my research assistant.
Effectively I’m using it like an automated search engine that indexes anything I want and refines the results by using the statistical near neighbors of how other people explained their searches.