Lambda the Ultimate

So, since I've learned about Kernel I was very excited: the idea of explicit evaluation seemed like a very cool idea, giving much more power to the programmer in comparison to the standard "pass evaluated arguments" strategy (1: this statement can be argued upon; 2: there were numerous posts here at LtU and other blogs about potential drawbacks).
Then, I've learned about Io language, which also seems to embrace the idea - when caller sends a message to a target, passed arguments are passed as expressions, not values, giving the full range of custom "control" messages, macros, etc.
This is when it hit me - although the idea sounds very cool, there is something wrong with it. Most likely my mind is stuck in an endless loop of dubious reasoning that I can't get out of, so, hopefully, someone can clarify my concerns.
Let us break down an example, where in some context we have:

a := Number(1) ;ignore how these two lines are actually executed
b := Number(3) ;whats important that the context has two number objects bound to symbols "a" and "b"
someAdderPrimitiveObject pleaseDoAddtheseNumbers(a, b) toString print

so, the caller context asks someAdderPrimitiveObject to add numbers a and b, and the arguments are just passed as "a" and "b" symbols. no problem here, as far as we concerned, because that same "someAdderPrimitive" object can ask the caller to send actual values back.
let's say we had defined the "pleaseDoAddtheseNumbers" something as

someAdderPrimitiveObject pleaseDoAddtheseNumbers := method(x, y, [body, whatever that is])

so, when the "pleaseDoAddtheseNumbers" method is invoked, the "a" and "b" symbols are bound in the environment of the "pleaseDoAddtheseNumbers" method's activation record to "x" and "y" symbols.
The method body would try and do something like this:

valx := caller pleaseEvaluateForMe(x)
valy := caller pleaseEvaluateForMe(y)
[do something with these values, whatever]

This is where it gets problematic for me. The callee (activation record of the "pleaseDoAddtheseNumbers" method) asks the caller (the original message sender) back for a value of its argument (which is bound to a locally known symbol "x") and in order to avoid infinite recursion of ping-pong of messages like "evaluate this for me", the callee *has* to pass the *value* of its own symbol "x" (bound to value, which is symbol "a") back to a caller, to ask it for a value (in this case: some boxed object-number 1).

So far as I've seen the problem is solved on an interpreter level, where this kind of thing is handled "behind the scenes".
Does that mean that the system that never evaluates passed arguments cannot implement itself, because at some point you *have* to pass values, in order for them to be operated upon?

Sorry if this is a mess, I hope someone undestands it :)