Autoresearch: The feedback loop behind self-improving agents

Autoresearch is an emerging approach to AI development that focuses on creating self-improving systems through feedback loops. Rather than relying solely on static models, this method uses an outer loop of agents to monitor, maintain, and refine the primary system. By utilizing feedback signals, evaluations, and human input, these agents can make architectural decisions and improve performance over time without requiring constant manual intervention.

A key component of this framework is the agent recipe, a portable format that captures the essential ingredients of a system’s evolution. These recipes document the models, evaluation criteria, and human expertise used to reach a specific state, allowing developers to track how past failures and decisions informed current capabilities. This process is designed to be extensible, using open-source building blocks like the Pi framework to ensure that systems remain portable and independent of specific model providers.

Human involvement remains a central part of these autonomous software factories. Initially, agents act like new employees, relying on human guidance to learn organizational preferences and workflows. As the system accumulates this tacit knowledge, it becomes increasingly autonomous. By treating the development environment as a miniature research laboratory, companies can build secure, proprietary agent systems that improve in efficiency and capability while keeping human expertise at the core of the production process.