Skip to content

Robin (FutureHouse)

external · status: active · focus: end-to-end · discipline: biomedical · started: 2025

Project page: https://github.com/Future-House/robin

Source: projects/landscape/robin.yml

Positioning

A multi-agent system for automating scientific discovery (arXiv:2505.13400), with explicit support for hypothesis generation, experiment design, and data analysis. The hypothesis and experiment modules are usable standalone; the full data-analysis path requires FutureHouse's commercial Edison platform.

Distinctive contribution

One of the few openly-documented RISE systems that explicitly separates hypothesis generation and experiment design as modular, runnable components rather than bundling them into a monolithic ideation step. FutureHouse-blog write-up demonstrates a full end-to-end scientific-discovery run.

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 2 Four stages spanning ideation through analysis; no drafting or review.
Autonomy level 2 Supervised: human researcher provides RQ; agents execute the discovery loop.
Architectural transparency 3 Open under Apache-2.0; arXiv paper + FutureHouse blog post; Docker setup published.
Inputs supported 2 Research-question inputs; integrated retrieval; multiple LLM providers via LiteLLM.
Outputs / reproducibility 1 Hypothesis/experiment paths runnable offline; full analysis path requires paid Edison API.
Internal evaluation 2 Demonstrated end-to-end run on a biological discovery task in the arXiv paper.
Openness 2 Source open under Apache-2.0, but full functionality gated behind commercial API credits.
Maturity / traction 2 301 stars; FutureHouse backing; recent (April 2026).
Cross-family policy 1 Edison platform agents + OpenAI/other via LiteLLM — cross-family supported.
Runtime assurance 2 Multi-agent literature → hypothesis → experiment loop with EdisonScientific platform validation.
Cross-platform portability 2 Docker setup + LiteLLM multi-provider; Edison API integration.

Scored on 2026-05-18. See the evaluation rubric.

Tags

Pipeline stages: hypothesis-generation research-design data-acquisition data-analysis

Architectural features: multi-agent tool-use rag-knowledge-base iterative-loop

Inputs: research-question

Outputs: hypotheses experiment-design analysis-results

Data sources: edison-platform literature

Knowledge sources: literature paper-qa

Limitations

  • Data-analysis path requires paid Edison API credits.
  • Biomedical orientation; portability to other empirical disciplines untested.
  • Quality depends on access to recent commercial LLMs.

Papers describing this project

  • Robin: A multi-agent system for automating scientific discovery — Ghareeb, A. E., Chang, B., Mitchener, L., Yiu, A., Szostkiewicz, C. J., Laurent, J. M., et al. (2025). arXiv. arXiv:2505.13400

Also compared in