Autonomous AI researcher that decomposes research objectives into experiments, assigns them to specialized GPU-enabled agents, and synthesizes findings into a paper-style writeup. It targets researchers and power users seeking to automate AI experimentation and accelerate discovery through a multi-agent, sandbox-based approach.

How It Works

The system employs an orchestrator to break down user prompts into discrete experiments. Specialist "researcher agents" are then assigned these tasks. Each agent can launch isolated, GPU-enabled sandboxes to perform model training, inference, evaluation, and evidence collection. The orchestrator aggregates results, iteratively refining the research direction or finalizing the output into a coherent research paper. This design leverages parallel, specialized computation via GPU sandboxes.

Quick Start & Requirements

• Primary Install/Run: Execute python run_app.py for an automated setup that installs dependencies, starts the API and frontend, and opens a web notebook.
• CLI Usage: Alternatively, create a virtual environment, activate it, run pip install -r requirements.txt, and then execute python main.py with desired arguments.
• Prerequisites: Requires API keys for LLMs and compute:
  • Google AI Studio: GOOGLE_API_KEY (for Gemini 3 Pro)
  • Anthropic: ANTHROPIC_API_KEY (for Claude Opus 4.5)
  • Modal: MODAL_TOKEN_ID and MODAL_TOKEN_SECRET (for GPU sandboxes) Keys can be configured via a .env file or entered interactively in the UI.
• Model Selection: Supports Gemini 3 Pro and Claude Opus 4.5, selectable via UI dropdown or the --model CLI argument.
• Deployment: Deployable via Railway by connecting the GitHub repository; a Dockerfile is detected for automated builds.
• Links: run_app.py opens a web notebook; no other specific documentation or demo links are provided in the README.

Highlighted Details

• Autonomous agent system for AI research tasks.
• Dynamic creation of GPU-enabled sandboxes for experiment execution.
• Generates structured, paper-style writeups of research findings.
• Supports leading LLMs like Gemini 3 Pro and Claude Opus 4.5.

Maintenance & Community

The project is explicitly described as a "super-early, experimental harness." No specific details regarding core maintainers, community channels (e.g., Discord, Slack), or a public roadmap are provided.

Licensing & Compatibility

The README does not specify a software license. This omission requires clarification regarding usage rights, modification permissions, and distribution terms, particularly for commercial applications.

Limitations & Caveats

This is a highly experimental project with known areas for improvement, including dataset sharing between agents, robust key management, and enhanced literature search capabilities. Its early-stage nature suggests potential instability and ongoing development.