Been messing around with autonomous agents recently (langchain, crewai etc) and the math on context growth is honestly terrifying.
because the agent has to re-read its entire past context on every single step, the token costs don't just grow linearly, they compound like crazy. one dumb infinite loop bug in your code while you're asleep and you wake up to a massive stripe charge from openai.
how do you guys handle this? do you just set hard spend limits on the api dashboard and pray? or is there a way to actually estimate how much a 50-step loop is going to burn before hitting run?
i'm honestly thinking about building a quick visual calculator to forecast this so i don't bankrupt myself, but curious how everyone else is managing the risk right now.