goalseek is a local-first Python package designed for running disciplined, git-backed research loops with coding-agent providers. The project is inspired by Karpathy's auto research project
It provides a structured environment for autonomous agents to propose changes, implement them, and verify results against defined metrics, while maintaining a complete audit trail via git and local artifacts.
The system is CLI-first but exposes its core functionality through a Python API, making it suitable for both interactive research and automated pipelines. The system also has a React Next JS UI to see the lifecycle of an an experiment
The project centers on the concept of a "Research Loop"—an iterative process where an agent (like Claude or Gemini or Codex or OpenCode or Client) is given a goal and a codebase, and then systematically attempts to improve a specific metric. Every change is tracked, every verification is logged, and every regression is automatically rolled back using git.
You can look at the system architecture details at this link
The Docusaurus documentation site is published at https://shambhu112.github.io/goalseek/.