### agent-agent signalingl it feels like this really

it feels like this really risks some kind of coding/hidden communication (esp if the same model is used for training itself)

e.g., in the tok sequence there is a pattern that triggers some circuit to accept the task