# The Agentic Scientist (AI-Scientist-v3)

> Open-source pipeline that orchestrates coding agents to turn a task description into experiments, ablations, and a written paper.

The Agentic Scientist, from Sakana AI, is an open-source pipeline that treats coding agents as AI scientists. Starting from a `task.md`, agents propose ideas, run experiments, evaluate results against a task rubric, and iterate, then the pipeline runs baselines, ablations, a sealed final evaluation, post-hoc analyses, and writes the paper.

## What It Is

Agentic Scientist is a command-line orchestration framework (`agsci`) for automated scientific discovery. It drives coding agent CLIs such as claude, codex, cursor, opencode, gemini, and pi through Headless CLI, so the same experiment can be repeated across different agents and models. Every agent phase runs inside Docker by default.

## How the Pipeline Works

Six phases run in sequence: task drafting, investigation, ablations, final evaluation, post-hoc analysis, and paper writing, with optional review and revision phases afterwards. Investigation is the idea, experiment, evaluate loop. Ablations compare the winning program against baselines and ablated variants, and the final evaluation re-scores it in a sealed environment. Each phase reads only a run-local snapshot of the task, so the shipped `examples/` are never modified.

## Outputs and Inspection

Each run writes a single directory under `results/<task>/<agent>/<run_id>/` containing drafting logs, the task snapshot, investigation records, ablations, final evaluation, post-hoc analyses, and the paper (manuscript, PDF, manifest, and optional AI reviews). Runs can be browsed in a local web UI with `agsci inspect start`. An OpenAI-backed reviewer can also be run on any PDF with `agsci review`.

## Setup Path and Safety

Setup involves building the Docker agent image, installing the Python environment with uv, and supplying agent API keys in a `.env` file. `agsci doctor` checks the host and `--dry-run` previews commands. The CPU-only circle-packing example is the recommended first run; a GRPO example requires GPUs. The README warns that agents run with full tool permissions and may inherit API keys, so runs should stay inside Docker and only trusted tasks should be used. Drafting a task from a bare research brief is described as experimental.

## Features
- Six-phase pipeline: task drafting, investigation, ablations, final evaluation, post-hoc analysis, paper
- Works with claude, codex, cursor, opencode, gemini, and pi agents via Headless CLI
- Docker-contained agent execution
- Sealed final evaluation and reproducible run-local task snapshots
- Baselines and ablation studies
- Automated paper writing and compilation with optional AI review
- Resume runs or run selected phases
- Run Inspector web UI
- Hydra key=value configuration and task presets
- GPU task presets
- Optional Autoillustrate concept figures

## Integrations
Claude, OpenAI Codex, Cursor, OpenCode, Gemini, Pi, Headless CLI, Docker, Autoillustrate, Amazon Bedrock

## Platforms
WINDOWS, LINUX, API, CLI

## Pricing
Open Source

## Links
- Website: https://github.com/SakanaAI/AI-Scientist-v3
- Documentation: https://github.com/SakanaAI/AI-Scientist-v3/blob/main/docs/pipeline.md
- Repository: https://github.com/SakanaAI/AI-Scientist-v3
- EveryDev.ai: https://www.everydev.ai/tools/agentic-scientist
