So, you're running into a couple of things I've been working around for about five years. Metrics that people aren't thinking about but should be.
Token diets are not the answer, because it's not the number of tokens that's causing the problem. It's the cyclomatic complexity of the operation, and the clarity of the instructions. A high number of tokens might not cause confusion, because the cyclomatic complexity is low. However, a small number of tokens with high cyclomatic complexity can.
Next up an observation. Have you ever noticed how sometimes, a 4b model can run circles around a 20b or a 60b? That's because it's not the number of parameters that's important. Here, the metric you want to think about is reasoning bandwidth, or recursive depth. How far can you take an agent, before it folds in on itself and loops?
You really have to approach both on a case by case basis. There is no hard rule that works on every system.
The room cleaning metaphor gives me an idea though. What if an agent had a tool where it could review its own context and delete anything it didn't feel like it needed to do the job? You could probably do it as part of your session management. Just revise between turns, the same way you would truncate in a custom client? I feel like that would work anywhere but Gpt-oss.
Oh well.
Just my thoughts.
Check the paper!
Love ya!