Agents in the Wild

FutureHouse's Robin agents proposed ripasudil as a dry AMD candidate; humans ran the lab tests

Robin, a FutureHouse multi-agent system, read the literature, proposed hypotheses and drug candidates, analyzed lab data and suggested follow-up experiments. It picked ripasudil, an existing glaucoma drug, as a candidate for dry AMD.

Organization
FutureHouse
Industry
AI Research
When
May 2025
Evidence
Research paper
Agents and tools
RobinCrowFalconFinchPaperQA2

Reported outcome

Robin's analysis agent reported ripasudil raised retinal cell phagocytosis 7.5-fold over controls in a lab assay; a human analysis of the same data found 1.75-fold.

  • 7.5-fold vs DMSO controls (paper)
    Phagocytosis increase (agent analysis)
  • 1.75-fold (paper)
    Same data, human analysis
  • 30; top 5 tested in round one, 10 in round two (paper)
    Drug candidates proposed
  • 151 for mechanisms, about 400 on RPE phagocytosis (paper)
    Papers reviewed

Caveats

Described in a paper with a stated method. Peer review status varies.

  • The agent's 7.5-fold result and the human's 1.75-fold result differ a lot on the same data. Check agent data analysis.
  • Humans did the wet-lab work. Robin did not run experiments.
  • This is a cell assay, not a clinical result. The authors say dose and timing tests are still needed.
  • The arXiv version is a preprint from the team that built the system.
  • The authors note the analysis agent depends on expert prompt engineering and its results vary between runs.

What the agents do

Robin combines literature agents (Crow and Falcon, built on PaperQA2) with a data analysis agent (Finch). It reviews papers, proposes disease mechanisms, ranks hypotheses with an LLM-judged tournament, suggests drug candidates and experiments, then analyzes the results and proposes the next experiment.

For dry age-related macular degeneration (dry AMD), Robin proposed boosting phagocytosis (cell clean-up) in the retinal pigment epithelium. It proposed 30 drug candidates. Humans tested the top ones in the lab. Robin then picked ripasudil, a drug already used for glaucoma in Japan, in a second round. It also proposed an RNA-seq experiment that pointed to ABCA1 as a possible mechanism.

What happened

Finch's analysis reported a 7.5-fold increase in phagocytosis versus controls, better than the reference compound Y-27632. A human analysis of the same data showed 1.75-fold. The authors say ripasudil still needs testing at different doses and longer times.

What to copy

  • Let agents do the reading and ranking; keep people on the experiments.
  • Run the analysis agent several times and compare. Robin runs 10 tries and takes a consensus.
  • Always have a human re-check the key number.

Starter prompt

Written by us, not by FutureHouse.

Research question: {question}
Papers (title, abstract): {papers}

1. List up to 10 possible mechanisms, each with 2 supporting papers.
2. For the top 3, propose one cheap lab experiment that could disprove it.
3. For each experiment, say what result would count as a failure.
Mark anything not directly supported by a paper as "speculation".

Sources

  1. Robin: A multi-agent system for automating scientific discovery — arXiv (FutureHouse, University of Oxford), May 19, 2025

Added Nov 2, 2026 · Sources last checked Nov 2, 2026. Spotted an error? Send a correction.

Get new use cases weekly

One short email. Sources included.

Related use cases