pi-autoresearch

pi-autoresearch is an extension for the pi AI coding agent that enables autonomous optimization loops by testing, benchmarking, and refining code changes. It supports various optimization targets such as test speed, bundle size, and LLM training metrics through a persistent session workflow. The tool includes a live dashboard, confidence scoring to filter out noise, and the ability to finalize experiments into clean, reviewable branches.

pi-autoresearch is an extension for the pi AI coding agent that facilitates autonomous optimization loops by iteratively testing, benchmarking, and refining code based on specific performance metrics. Its architecture separates domain-agnostic infrastructure from domain-specific skills, allowing the agent to maintain state across restarts via persistent log files and session documentation. This design solves the problem of manual, repetitive benchmarking by automating the 'try-measure-keep' cycle, though it requires careful management of API token usage and benchmark noise to ensure reliable results.