I talked my local, qwen3.6-based pi coding agent into better managing contexts (lean_ctx seems pretty nifty, as does anvil for us Emacs-philes). My tk/s dropped in half. QUALITATIVELY I'm receiving better responses more quickly. I don't feel like running a full experiment on myself, and only one user wouldn't be very meaningful anyways. My productive use shot upward even with a tokens-per-second drop.

Methinks the proper, problem-to-result measurements are not yet established and should not be used to evaluate these situations.