The authors propose TAHI, a framework that adapts AI agents to individual users by leveraging interaction data accumulated across multiple task sessions. In experiments with 30 individuals on writing and visual creation tasks (600 total), agents improved solo task success by 4.5-20.9% after only tens of tasks. Additionally, an evolving rubric module derived from these interactions catches 16.0-22.3% more failures than rubrics from language models or humans alone, and personalized agents show up to 8.8% improvement that generalizes to other users.
This work is interesting because it addresses the gap between general AI capabilities and the specific, often unverbalized criteria that individual experts require. By reusing the natural signals from iterative human-agent collaboration—such as edits, feedback, and plan adjustments—instead of treating each session independently, the method achieves efficient personalization with minimal user burden.