Signals: Toward a Self-Improving Agent

Signals is an innovative system designed to enhance traditional product analytics by evaluating user experiences through large language models (LLMs), identifying moments of friction and delight that standard metrics often overlook. By analyzing user sessions without revealing sensitive information, Signals extracts abstract patterns and categorizes them into facets and friction types, allowing for a nuanced understanding of user interactions. This system operates at scale, processing thousands of sessions daily, and provides insights into user behavior by correlating these patterns with backend system logs and release data. Signals goes beyond identifying problems, aiming for recursive self-improvement by autonomously suggesting and implementing fixes, and it continuously evolves by detecting emerging patterns such as context churn and specification drift. The ultimate goal of Signals is to create a self-evolving agent capable of real-time adjustments and proactive development by learning from user interactions and identifying opportunities for new capabilities.