agr

# AGR: Artificial General Research > Autonomous iterative optimization for any measurable problem. > Powered by **Claude Code**. Inspired by [Karpathy's autoresearch](https://github.com/karpathy/autoresearch). > > Thank you Andrej Karpathy for everything you do for open source and the AI community. > Your autoresearch showed that AI agents can run experiments overnight and wake up to > a better system. AGR generalizes that vision beyond ML to any domain. > > Also inspired by [uditgoenka/autoresearch](https://github.com/uditgoenka/autoresearch) > for the Guard/Metric separation and rework-on-failure patterns. > > **Requires**: [Claude Code](https://claude.com/claude-code) v2.1.72+ (`claude` CLI in PATH) ## What Is AGR? AGR uses Claude Code's headless mode (`claude -p`) to run an autonomous optimization loop. Each iteration is a **fresh Claude Code instance** (no context degradation) that: 1. Reads state from files (what's been tried, what worked, what's next) 2. Picks ONE experiment to try 3. Implements the change 4. Measures **Metric** (the number to optimize) + checks **Guard** (must not break) 5. Keeps if improved, reworks if guard failed, discards if no improvement 6. Logs everything, updates strategy document 7. Dies — next iteration starts with fresh context **Works for any problem where you can**: measure a number + verify correctness + modify code. ## Quick Start ```bash /agr speed # set up speed optimization /agr accuracy # se...