this harness engineering paper is f*cking insane
it's the exact setup anthropic used to keep a single AI agent coding for hours without forgetting what it was doing.
2 agents, 4 files, one prompt that runs every session, all on one page.
set it up tonight and your agent keeps building feature after feature while you sleep, instead of quitting halfway.
the crazy part is everyone blames the model when their agent gives up after 20 minutes, while the fix is a json file and a text log.
full breakdown in the harness guide below.