A team runs an agent that builds a large application over many hours, across many separate context windows. Each new session starts with no memory of the last, and the agent tends to redo finished work, declare the project done early, or leave the code half-broken. Which harness design best addresses this?