pi-autoresearch, an AI coding agent with an autonomous optimization loop

pi-autoresearch extends the capabilities of pi, a terminal-based AI coding agent, providing an autonomous optimization loop that automates the process from applying ideas to measuring benchmarks and adopting or reverting results.

Key features are as follows:

• Autonomous experimentation loop: Repeats the process of modifying code, running benchmarks, recording results, and deciding whether to adopt or revert based on performance improvement, without human intervention.
• Domain-agnostic design: Designed to be applicable regardless of what is being measured, whether it's test execution time, bundle size, LLM training metrics (val_bpb), build time, and more.
• Separation of Extension and Skill: Increases versatility by separating the extension feature, which provides common tools, from the skill, which contains domain-specific knowledge (commands, metrics, scope, etc.).
• Sustainable workflow: Records experiment results in files and dashboards within the project, allowing work to continue while maintaining context even if the session restarts.