FAROS: Research Automation as a Runtime, Not a Prompt Stack

What it does FAROS is a workflow runtime that automates LLM-domain research from ideation through paper generation and simulated peer review. You bind a blueprint—currently just the ml_paper workflow—to a profile like faros_llm, and the runtime orchestrates idea refinement, experiment scaffolding, venue-aware LaTeX drafting, and reviewer simulation. It persists runs, events, and artifacts to disk, and exposes the whole state through a REST API.

The interesting bit Instead of hardcoding another end-to-end AI scientist, FAROS separates workflows (Blueprints), execution steps (Capabilities), runtime strategies (Profiles), and backends (Providers). The current release is essentially a proof-of-concept for a broader vision where future domains plug in non-LLM providers without touching the orchestration core. It also generates real conference-formatted LaTeX—ICML, NeurIPS, ICLR, ACL—and falls back to a simplified PDF if latexmk chokes.

Key highlights

• Blueprint-driven orchestration: declarative workflow graphs with constraints and validation, not a frozen agent pipeline
• Capability adapters wrap existing native modules (idea, code, paper, review, platform) rather than replacing them
• Venue-aware paper generation with bundled LaTeX templates and a PDF fallback when compilation fails
• File-backed persistence for research memory, artifacts, and run history—no database required yet
• Plan-only execution mode lets you dry-run a workflow before triggering full LLM-backed generation

Caveats

• Explicitly a release candidate (1.1.0-rc1) scoped to the LLM domain; full experiment execution, DAG scheduling, parallel orchestration, and cross-domain providers are listed as not yet included
• Persistence is file-backed, the frontend console is unfinished, and the “full experiment execution and evaluation loop” is still missing
• Cross-domain abstraction remains aspirational: the README stresses that generalization comes only after the first workflow is coherent

Verdict Worth exploring if you are building agentic research tools and want a modular runtime architecture rather than a monolithic prompt chain. Skip it if you need a mature, fully automated experimental evaluation loop or immediate support for non-LLM scientific domains.