Skip to content

Getting Started with ARI

ARI is an end-to-end autonomous research system: give it a plain-text research goal and it surveys prior work, forms a hypothesis, runs real experiments, writes a paper, and verifies its own reproducibility. This page is the map for your first hour.

A learning path

Follow these in order — each step assumes the previous one.

  1. Quickstart — install ARI, pick an AI model, and launch your first experiment from the web dashboard. Operation-focused: which button does what.
  2. Your first experiment, end to end — a narrative walkthrough of a single small experiment from goal to reproduced paper, explaining why each stage exists. Read this once the dashboard works.
  3. FAQ — the questions newcomers hit first: model choice, the 8765 port, where output goes, GPU/SLURM detection, "why do my child nodes show the same numbers?".
  4. Glossary — one-line definitions of the recurring terms (BFTS, frontier, rubric, venue, EAR, ORS, CoW, …) so the concept docs read smoothly.

Then branch by what you need

If you want to…Go to
Write a good experiment.mdWriting experiment files
Run on a SLURM/HPC clusterHPC setup
Understand how the search worksBFTS algorithm · Architecture
Reproduce or audit a published paperPaperBench quickstart
Add your own capability (skill)Extension guide
Drive everything from the CLICLI reference
Fix something that brokeTroubleshooting

Two things worth knowing up front

  • The dashboard runs on port 8765. Start every service with ./start.sh at the repo root and open http://localhost:8765; stop with ./shutdown.sh.
  • Each run is self-contained. All state for a run lives under workspace/checkpoints/<timestamp>_<slug>/ — nothing is written to your home directory, and API keys come from .env, never from the saved settings.

See also: Quickstart · FAQ · Glossary · Documentation index