Robin (FutureHouse)

First multi-agent system to integrate hypothesis generation with experimental data analysis in a lab-in-the-loop workflow. Built on the Aviary framework, validated wet-lab in dry age-related macular degeneration.

   
Affiliation FutureHouse, with collaborators at the University of Oxford and Fordham University (FutureHouse)
First introduced 2025-05 (preprint); peer-reviewed paper accepted 2026-05
Lifecycle stages Multi-stage (hypothesis generation + experiment design + analysis; lab-in-the-loop)
Autonomy level Semi-autonomous (closed-loop with checkpoints; humans run wet-lab experiments)
Domain focus Biology / therapeutics
Availability Open source (code) + closed agent platform (FutureHouse-hosted Crow / Falcon / Finch endpoints)

Approach

Multi-agent Jupyter notebook built on the FutureHouse Aviary framework. Combines PaperQA2-based literature search agents (Crow for concise, Falcon for deep summaries) with Finch, a Jupyter-native data analysis agent that runs analyses across multiple independent trajectories and produces consensus conclusions.

Validation

Wet-lab study on dry age-related macular degeneration (dAMD). Robin proposed enhancement of retinal pigment epithelium phagocytosis as a strategy, selected drug candidates, analyzed RNA-seq follow-up, and produced all hypotheses, analyses, and figures in the main text.

Notable results

  • Identified ripasudil (a clinically used ROCK inhibitor) as a novel dAMD candidate, confirmed in vitro.
  • Identified KL001 as a second hit.
  • Proposed and analyzed an RNA-seq follow-up that revealed ABCA1 upregulation as a candidate mechanism/target.

Primary paper

Ghareeb et al., “A multi-agent system for automating scientific discovery,” Nature 2026.

Other references

Code