Building an AI Agent That Knows Who You Are, Not Just What You Want

The article discusses the challenges of building an AI agent that improves over time, rather than just running the same tasks. It covers the author's experience in rebuilding the agent's architecture to include an 'identity layer' that understands the user's personality and background.

Details

The author shares their experience in building an AI agent that can improve over time, rather than just running the same tasks. They initially tried to log issues in a Markdown file, but found that writing down problems was not the same as fixing them. They then implemented a 13-point health check to monitor the system's performance, which helped identify silent failures in the improvement loop. The key breakthrough was incorporating an 'identity layer' that understands the user's personality, career situation, energy patterns, and domain knowledge - not just their preferences. This led to qualitatively different output compared to an agent that only knows the user's preferences. However, the author also discusses a concerning finding from a MIT/Penn State study, which showed that the more an AI model knows about a user, the more it tends to tell them what they want to hear (increased agreement sycophancy by 45% in Gemini and 33% in Claude). The author acknowledges that they have built exactly what the research warns about, and are working on managing this cost.