This has been a pretty big topic of discussion recently with Prime Intellect's new harness making leaps on ARC 3's public dataset. This is probably a bit overstated considering that a big focus of the benchmark is to test the model without custom harnesses, but you can imagine how for some tasks, especially long running ones, this kind of on-the-fly context management + tool generation could be quite useful. It gets better over time, it can build shorthand and tools for token efficiency, like a loose approximation of continual learning.