# 04 — FutureHouse’s Robin & Kosmos: Closed-Loop Discovery and Structured World Models **FutureHouse** (founded in 2023 by Sam Rodriques and Andrew White, backed by Eric Schmidt) and its commercial spinout **Edison Scientific** tackled two fundamental engineering bottlenecks that break standard research agents: 1. **Robin (*Nature*, 2026):** Closing the full *"literature $\rightarrow$ hypothesis $\rightarrow$ experiment $\rightarrow$ data analysis $\rightarrow$ next hypothesis"* loop in a real wet lab. 2. **Kosmos (*arXiv*, late 2025):** Solving **long-horizon context decay** so an AI scientist can stay coherent across 12-hour runs reading 1,500 papers and executing 42,000 lines of analysis code. --- ## 1. Robin: First End-to-End "Lab-in-the-Loop" Multi-Agent System * **Citation:** Ghareeb, A. E., Chang, B., Mitchener, L., Yiu, A., Szostkiewicz, C. J., Shved, D., Gyimesi, G. J., Laurent, J. M., Wright, S. M., Razzak, M. T., White, A. D., Finnemann, S. C., Hinks, M. M., & Rodriques, S. G. *"A multi-agent system for automating scientific discovery."* ***Nature*** **655**, 497–505 (Published online May 19, 2026; Print July 2026), DOI: `10.1038/s41586-026-10652-y` (Preprint: `arXiv:2505.13400`). * **Institutions:** FutureHouse, University of Oxford, Fordham University. ### Architecture: Specialized Literature & Code Agents Robin coordinates three specialized sub-agents built atop FutureHouse's **PaperQA2** engine (an agentic full-text PDF/citation-graph RAG system that beat human PhDs on FutureHouse's `LitQA2` benchmark): * **Crow:** Fast, concise literature-orientation agent. * **Falcon:** Deep-dive literature synthesis agent that writes candidate mechanism reports by reading full papers and chasing citations. * **Finch:** Autonomous data-analysis agent that writes and executes Python/R scripts to process wet-lab readouts (such as RNA-seq gene expression matrices and flow cytometry cell counts), updating Robin's hypotheses in the next cycle. ### Real-World Wet-Lab Discovery: Dry Age-Related Macular Degeneration (dAMD) * **Biology Context:** Dry AMD is a leading cause of irreversible blindness in older adults. In healthy eyes, support cells called **retinal pigment epithelium (RPE)** act like garbage collectors (`phagocytosis`), constantly clearing waste from light-sensing photoreceptors. In dry AMD, RPE waste clearance breaks down. There are very few effective drugs. * **Autonomous Hypothesis Generation:** Given dry AMD as a target disease, Robin read across ophthalmology, cell biology, and pharmacology literature and independently proposed that **enhancing RPE phagocytosis** could treat the disease—and nominated **ripasudil** (an approved glaucoma Rho-kinase / ROCK inhibitor) and **KL001** (a circadian-clock modulator), **neither of which had ever been proposed for dry AMD**. * **Closed-Loop Wet-Lab Validation:** Human lab scientists executed Robin's proposed experiments on stem-cell-derived RPE cells and fed the resulting **RNA-seq** datasets back into Robin's **Finch** agent. * **Mechanistic Insight:** Analyzing the experimental transcriptomics, Robin confirmed the functional boost and traced the molecular mechanism to upregulation of **`ABCA1`**—a lipid efflux pump gene that helps RPE cells clear cholesterol/waste. --- ## 2. Kosmos: Scaling Long-Horizon Science via a Structured World Model * **Citation:** Mitchener, L., Yiu, A., Chang, B., ..., Ghareeb, A. E., White, A. D., & Rodriques, S. G. *"Kosmos: An AI Scientist for Autonomous Discovery."* ***arXiv:2511.02824*** (November 2025). * **Institution:** FutureHouse / Edison Scientific. ### The Engineering Problem Kosmos Solves Standard LLM agents fall apart after 10–30 tool calls: their context window fills with noisy stack traces and redundant paper snippets, causing them to lose track of what they already tested (**"horizon collapse"**). ### Architecture: Parallel Agents + Shared World Model Instead of passing a giant linear chat transcript around, Kosmos separates **execution workers** from **persistent scientific state**: * **Structured World Model:** A central, continually updated knowledge graph / structured state store recording verified facts, open hypotheses, failed analyses, code artifact pointers, and literature citations. * **Parallel Data-Analysis & Literature Agents:** Dozens of parallel trajectories branch off to write analysis scripts or read full-text papers, writing *only verified summaries + provenance links* back into the shared World Model. * **Strict Provenance Contract:** Every single claim in Kosmos's final report must explicitly point to either **(a)** a specific passage in a cited primary paper or **(b)** the exact executable Jupyter notebook / line of code and dataset output that produced it. ### Scale & Verification Metrics * **Single Run Scale (`~12 hours`, `20 cycles` / `~200 agent rollouts`):** * Reads approximately **`1,500` full-length scientific papers**. * Writes and executes approximately **`42,000` lines of analysis code**. * **Independent Accuracy Audit (`79.4%`):** When independent domain scientists audited statements in Kosmos's generated reports line-by-line, **`79.4%` of specific claims were verified as accurate**, with beta collaborators estimating that a single 12-hour Kosmos run compressed roughly **6 months of PhD/postdoc-level data exploration into a day**. ### Commercialization & Pharma Adoption (2026) Demand for Kosmos led FutureHouse to launch a commercial spinout, **Edison Scientific** (building the Edison AI Platform): * **Incyte Collaboration (May 2026):** Embedded Kosmos across translational and clinical discovery datasets. * **Pfizer Collaboration (Late 2026):** Deployed Kosmos across Pfizer's R&D pipelines in oncology, immunology, and internal medicine to build compounding internal world models on proprietary pharma data. --- ## 3. Key Engineering Takeaways * **Externalized World Models > Long Context Windows:** Even with 1M-token LLMs, scientific agents need an **explicit, structured shared state (World Model)** so 200 parallel rollouts don't drift or repeat work. * **Code-Traceable Citations:** In biology, a bare URL citation isn't enough—agents must link every numerical claim directly to the **exact script and statistical output** (`Finch` / `Kosmos`) so human scientists can audit whether the agent cherry-picked a $p$-value or mishandled normalization.