.@danshipper on the agent death spiral that burns millions of tokens. Most AI models are trained for one-on-one conversations, so they don't quite know the etiquette of a group chat—like when to stay quiet. Left unchecked, this can trigger something similar to a phenomenon in nature known as the ant mill. Blind army ants, following the pheromone trail of the ant in front of them, can end up walking in an endless circle—with each ant laying fresh scent that keeps the next ant following—until they die of exhaustion. A similar thing can happen with OpenClaw: if one agent messages a channel that other claws are in, and the settings aren’t right, the agents will go back and forth endlessly—burning millions of tokens in the process—until a human steps in and pulls the brakes. Getting agents to collaborate fruitfully requires a fundamental shift at the model layer. Link to the full episode on what happened at Every when every team member got an agent: