Reward Models and Utility Functions (Without the Math Pain) in Agentic AI
Reward Models and Utility Functions (Without the Math Pain)
Utility in human terms
Utility is just a score for outcomes. You decide what matters: correctness, user satisfaction, time, cost.
A simple utility template
score = (quality * w1) - (cost * w2) - (risk * w3)
Why explicit utility helps
If you don’t define utility, the agent will implicitly optimize for “sounds helpful”, which is not always correct or safe.

