FAROS (Foundation AutoResearch Operating System) is a multi-agent system built around the scientific workflow. It connects research ideation, literature retrieval, planning, code generation, experiment tracking, paper writing, and peer review, and keeps provenance, decisions, artifacts, and human feedback at every stage.
The pipeline runs from a research interest to an evidence-grounded idea, a validated PlanPackage, executable code, a measured experiment, a traceable paper, a ReviewX consistency audit, and then human approval or an evidence-driven revision. PlanPackage is the typed contract between modules: it freezes hypotheses, variables, steps, acceptance criteria, and evidence references so a downstream module does not have to infer upstream intent. Retrieval, semantic alignment, deep reading, and evidence gates run before ideas are generated, and ReviewX locates problems at the claim, evidence, or measurement level and emits revision actions.
The runtime follows a Blueprint + Capability + Profile + Provider design so new workflows, tools, and scientific domains can be added without rebuilding the orchestration core. The backend is a FastAPI service (Python 3.11+) with a React and TypeScript workbench (Node.js 18+); Docker Engine is optional for local browsing and required for the production code and experiment sandbox, and formal paper PDFs need latexmk, XeLaTeX, the ctex package, and a CJK font. At least one compatible LLM provider is required, with Qwen recommended, and provider credentials are encrypted and isolated per user.
The current release is a release candidate aimed at competition validation and research prototyping rather than unsupervised research. Cross-domain experiment execution, additional domain blueprints, large-scale parallel scheduling, and broader human evaluation are listed as active work, and the project states that important scientific claims should be checked against their original evidence. The published validation baseline is 644 backend tests and 35 frontend tests passing, along with a TypeScript production build.
Features
- Idea stage: topic coaching, literature retrieval, semantic filtering, deep reading, and novelty and feasibility review
- PlanPackage: hypothesis refinement, variables, staged execution, a reviewer committee, and human approval
- Code stage: code retrieval, project generation, sandbox execution, static and dynamic evaluation, and repair
- Experiment stage: metric ingestion, dataset upload, result comparison, and publication figures
- Paper stage: brief, outline, section drafting, citation management, and LaTeX or PDF generation
- ReviewX: claim extraction, evidence alignment, consistency and reliability checks, and a human feedback loop
- Human gates: people approve plans, interpret results, and sign off, with feedback kept in versioned records
- Recoverable long jobs: background tasks and polling survive a browser disconnect without duplicate model calls
- Deployment isolation: bilingual interface, light and dark themes, per-user provider configuration, and encrypted API keys
