How AI research agents are being built to sustain long-running investigations, carry memory across sessions, learn from their own experiments, and even rewrite their own research code, plus how we can tell whether those runs actually reach sound conclusions.