For years AI in the lab meant a better assistant — summarize this, retrieve that, run this pipeline. The 2025 frontier is different: multi-agent systems that generate their own scientific hypotheses, debate them against each other, and evolve the survivors. Google’s AI co-scientist runs an Elo-scored ‘idea tournament’ to surface novel, wet-lab-confirmed hypotheses; FutureHouse’s Robin went further and closed the loop — proposing, testing, and interpreting its way to a genuinely new drug candidate for a leading cause of blindness, with humans only running the pipettes. It’s the reasoning layer above every model we cover, and it’s exactly the question the lab is asking about its own agents.
