Will Robertson wrote:
>> Does it inspect the argument at runtime to figure out which of the two
>> cases is at hand? Looks rather fragile & syntactic sugar to me.
> 
> The heuristic is fairly safe, I believe. It uses \ifx\noexpand#1#1 to
> detect whether the argument is a macro, and if it is not inserts \the to
> extract the value of the toks/dim/whatever. The whole thing takes place
> inside a \romannumeral expansion so the number of expansion steps is
> constant.

Thus far, all seems to work okay. Whether that means it never goes wrong
 ...

> (There is no error checking at present, but may be added conditionally
> in the future.)

Will, don't forget that it does catch things like:

\csname l_does_not_exist_tl\endcsname

which using :o will not. (Variables are not meant to equal \relax under
any circumstances, according to the LaTeX3 specs.)

>>> So V is preferred for accessing variables (as the internal structure is
>>> not important),
>>
>> Wouldn't you still need to know the variable's "internal structure"
>> when setting it?
> 
> But setting does not need to be done in a known number of expansion
> steps, so the setting functions \toks_set:Nn, \tl_set:Nn, etc., all
> behave analogously.

As Will says, we have set functions for everything, and also use
functions, the idea begin that you can always know that

\<var>_use:N \l_my_var

and

\<var>_set:Nn \l_my_var {<content>}

will be correct (of coure <content> must be correct here!).
-- 
Joseph Wright