Robin

Robin is a multi-agent system from FutureHouse, described in a Nature paper titled "A multi-agent system for automating scientific discovery". It combines literature search agents with data analysis agents in a lab-in-the-loop workflow. The system generates hypotheses, proposes experiments, interprets experimental results and produces updated hypotheses. Wet-lab work and some steps, such as read demultiplexing and alignment, are performed by humans.

The component agents are Crow (literature review), Falcon (deep reports) and Finch (data analysis). In the paper, Robin was applied to dry age-related macular degeneration. It proposed enhancing retinal pigment epithelium phagocytosis as a strategy. It identified ripasudil, a Rho kinase inhibitor, and KL001 as candidates with in vitro efficacy. A follow-up RNA-seq experiment analysed by Finch showed upregulation of ABCA1, which the authors suggest is a possible novel target. A YouTube creator reported figures of under 2 hours and about $10.76 of compute, against an estimated 400 human hours. These figures are not confirmed in the provided paper text.

The authors state current limits. Robin produces experimental outlines rather than precise executable protocols. Finch relies on prompt engineering. Results are in vitro only. The paper refers to a Robin GitHub repository with the code, including prompts. The URL was not in the provided material. Access is mainly through the paper and the released code, aimed at researchers.