Agent adapted to a user boosts solo success to 20.9%

Researchers adapted an agent to 30 people across 600 tasks and raised solo success 4.5-20.9% within tens of tasks, per TAHI. That's the difference between an agent that stays generic and one that adapts fast enough to serve a single new user well.

TAHI folds a user's past sessions into both the agent's working context and its weights. A rubric module distills what a person wants from a task, then scores new attempts against that rubric. The 600 tasks spanned writing and visual creation, work where 'good' depends on who's asking.

The same rubric module doubles as a standalone annotation tool. Its rubrics catch 16.0-22.3% more failures than rubrics written by models alone or by humans alone, with no personalization step required.

The paper doesn't say how the gains hold up past 600 tasks or past the 30 people tested. It's unclear whether improvement keeps climbing as a user's history grows into the thousands.