pi-autoresearch is an extension for pi, an AI coding agent, that enables autonomous optimization loops.
The tool solves the problem of manually iterating on code improvements by automating the cycle of trying ideas, measuring results, and keeping what works. It provides three core tools: init_experiment for configuring a session with a metric and optimization direction, run_experiment for executing commands and capturing timing and output, and log_experiment for recording results and auto-committing changes. The extension maintains a live dashboard and stores experiment history in a log file, allowing the agent to resume optimization loops across sessions.
pi-autoresearch suits projects where you want to continuously optimize a measurable target—whether test speed, bundle size, build times, or other metrics. It works best when integrated into pi's workflow as an extension rather than as a standalone tool. The project is inspired by karpathy/autoresearch and applies that pattern to any optimization goal. Users can interact with the tool through subcommands like /autoresearch to enter optimization mode, /autoresearch export to view a live browser dashboard, or /autoresearch dashboard for a terminal overlay. Keyboard shortcuts are optional and configurable to avoid conflicts with pi's built-in keybindings.
Development activity shows a focused, deliberate approach to design. The project avoids binding default keyboard shortcuts, instead requiring explicit opt-in configuration to prevent collisions with pi's evolving keybindings. Documentation is comprehensive, with a dedicated website covering configuration and usage. The extension is tightly scoped to the autonomous loop workflow, providing only the tools needed for experiment management rather than attempting broader functionality.
