Chapter Assessment: A Research Agent

This review covers the second complete design: the upgrade research agent, its tasks fixed by a frozen snapshot, its claim-level graders, its repeated runs and cost, its planted attacks, and its eval plan. The bank has 20 questions, and each attempt draws ten, so repeated attempts reach different parts of the design. The weight falls on decisions about one agent: which failure mode a trace shows, which measure a claim fails, and whether a result clears a gate