autoresearch

This repository is a bootstrap for an autonomous LLM agent that is expected to run AI research end to end with minimal human intervention.

The repo includes several instruction templates:

File names are additive: a name includes the features that are present in that template.

• FROM-SCRATCH: starts without paper or existing-codebase inputs
• BUILDING-ON-TOP-OF: includes paper and existing-codebase inputs
• DRIVE: includes remote storage guidance via rclone
• HARDWARE: includes local compute or hardware guidance

Available templates:

• AGENTS-FROM-SCRATCH.md: the base from-scratch template
• AGENTS-FROM-SCRATCH-DRIVE.md: from-scratch template with drive guidance
• AGENTS-FROM-SCRATCH-HARDWARE.md: from-scratch template with hardware guidance
• AGENTS-FROM-SCRATCH-DRIVE-HARDWARE.md: from-scratch template with both drive and hardware guidance
• AGENTS-BUILDING-ON-TOP-OF.md: paper/codebase variant
• AGENTS-BUILDING-ON-TOP-OF-DRIVE.md: paper/codebase variant with drive guidance
• AGENTS-BUILDING-ON-TOP-OF-HARDWARE.md: paper/codebase variant with hardware guidance
• AGENTS-BUILDING-ON-TOP-OF-DRIVE-HARDWARE.md: paper/codebase variant with both drive and hardware guidance

The core behavior defined by these templates includes:

• search for relevant libraries, repos, and techniques before building custom solutions
• set up the environment and dependencies explicitly
• when the chosen template includes them, use the available storage and hardware resources for the project
• implement iteratively with tests, logging, and regular execution checks
• run experiments, inspect outputs, record findings, and revise the plan continuously
• keep README.md and AGENTS.md up to date as living project documents
• commit and push progress regularly to GitHub
• produce end artifacts including code, experiment results, model outputs, and a LaTeX research paper

The intent is not just to write code. The agent is instructed to own the full loop of AI research and engineering work: planning, implementation, experimentation, debugging, documentation, and paper writing.

In practice, this repo serves as an execution scaffold for autonomous AI research projects where an LLM agent is expected to:

• work through an AI research problem from setup to results
• make grounded decisions based on code, experiments, and external research
• keep an explicit running plan in version control
• validate that results are meaningful, reproducible, and well logged
• continue iterating until the project reaches a complete research-paper stage

The workflow assumes uv for environment management and GitHub for version control. Some variants also include optional remote artifact storage via rclone, for example rclone ls googledrive:projects/<project-name>.

If you want to understand how the agent should behave, start with the template that matches your project constraints.

Do not remove any lines above the Scratchpad.

Scratchpad