There is a long history of distinguishing external arguments that seem to be volitional agents from both inanimate causers and instruments. Concentrating on Alexiadou and Schafer (2006) here (see their paper for more references), I show that the arguments for distinguishing them do not go through. The alternative is to say that all of them are the same, and are distinguished only by world knowledge. Alexiadou and Schafer call this view the "underspecification" view of external arguments. Borrowing the term from Ramchand (2008), let us refer to this underspecified external argument role as "Initiator." So the two views we are contrasting are the following:
A. All external arguments are Initiators.
B. External arguments must be divided into agents, causers, and instruments, which behave differently.
Alexiadou and Schafer (2006) present several arguments against a version of the Initiator view. However, the version they argue against is so simplistic as to be unworkable even without arguing against it. Their arguments against such a theory have no force against a more sophisticated version. Here is the crucial addition:
Every theory has to allow semantic selection for properties of the subject. Consider the following contrast:
(1) A rampaging elephant killed the Prime Minister.
(2)#A rampaging elephant murdered/assassinated the Prime Minister.
Unlike kill, murder and assassinate require subjects that are capable not just of volitional action, but also awareness of morality and politics. Similarly, the verb eat requires a subject that is capable of ingesting:
(3) The sea witch ate the captain's heart.
(4)#The sea ate the beach. (vs. The sea ate away at the beach, no such requirement)
These selectional requirements are over and above the distinction between agents and causers (or instruments). Every theory that I am aware of would treat the subjects of eat, kill, murder, and assassinate as agents.
So, every theory requires semantic selection in addition to whatever thematic roles it assigns to external arguments. In the Initiator theory, then, external arguments can all be assigned the broad role of Initiator, but individual verbs (or prepositions, or VPs) might impose additional selectional restrictions. With this in mind, let us examine Alexiadou and Schafer's arguments for distinguishing agents from causers.
Their first argument is that Greek passives only allow animate agents and not inanimate causers. So the Greek equivalent of The clothes were dried by the sun is reported to be ungrammatical (their example 2b). Similarly, some languages, like Jacaltec, are reported to only allow animate subjects of transitive verbs. Conversely, certain prepositions only allow causers:
(5) The window cracked from the pressure / *from Will.
Although one would want to know much more about Jacaltec, it is very easy for the Initiator theory to account for it: through selectional restrictions. Suppose transitive Voice in Jacaltec uniformly selects only animates; then the facts are accounted for, without any need to refer to a role of "agent." Similarly, the preposition from in the relevant sense in English selects nouns that denote events. Hence nouns like pressure, dancing, thermal expansion but not Will are good complements of from.
As for Greek, Alexiadou and Schafer go on to show that, in fact, inanimates are allowed in Greek passives; their examples in (42b-c) have The door was opened by THIS key/ the ELECTRONIC key (with focus intonation). I will return to this momentarily.
The last argument Alexiadou and Schafer give against the Initiator account is that, according to them, it cannot account for why some apparent instruments make good subjects while others do not. In fact, though, it accounts for the facts straightforwardly.
Here is the basic contrast:
(6) John loaded the truck with a crane/pitchfork.
(7) The crane/*pitchfork loaded the truck.
The Initiator theory accounts for this easily: external arguments have to be capable of being Initiators. That is, they have to be able to initiate events. In our conception of the world, cranes can do that, because they are self-propelled and capable of motion. Pitchforks are not. Change it to My new autopitchfork and (7) becomes fine. The same account explains the entire range of data they give in English, German, Dutch, and Greek, shown here only for English:
(8) The crane picked the crate up.
(9)#The fork picked the potato up.
(10) The falling axe broke the windowpane.
(11) The storm broke the windowpane.
(12) The chamomile cured Mary.
(13)#The scalpel cured Mary.
Again, cranes are capable of independent motion, whereas forks are not. (However, I personally find fork fine with modals or focus: "These plastic forks are just going to break. Only this titanium fork can pick up such a heavy potato." Again, see below.) Falling axes can initiate events by virtue of their motion; so can storms. Chamomile can because of its chemical properties. Scalpels, on the other hand, cannot do anything on their own. (Again, though, a scalpel is fine as a subject in certain circumstances: "THIS scalpel can rid/cure Mary of her unsightly blemishes!")
Alexiadou and Schafer also make an unmotivated distinction between "tools" and "secondary tools" (following Nilsen 1973). This distinction is supposed to be the following (and note that modals and focus do not help with 16b and 17b):
(14)a. Ashley cut the melon with a knife.
b. This knife cuts melon easily.
(15)a. Casey opened the door with the key.
b. This key opened that door.
(16)a. Cathryn ate spaghetti with a fork.
b.#This fork ate spaghetti.
(17)a. Denis is drinking juice with a straw.
b.#This straw is drinking juice.
There is no independent way of determining whether a tool is a "tool" or a "secondary tool." Moreover, the distinction is a spurious one; as shown above, the verbs eat and drink require external arguments that are capable of ingesting. It is the verbs that are behind the ill-formedness of (16b) and (17b), not the particular tool involved. Changing the verb changes the judgments:
(18)a. Cathryn scooped up spaghetti with a fork.
b. This fork can scoop up lots of spaghetti.
So, why do modals or focus help to make inanimates better external arguments? I submit that it is because they focus on inherent properties of the inanimate entity, properties that make it capable of carrying out the action of the verb. For instance, in (18b), using the ability modal focuses on the abilities of the fork, namely its tensile strength and scooping capacity (and retention, without spillage). These abilities, once they are brought to the conceptual fore, make the fork a good Initiator.
Finally, a major problem for Alexiadou and Schafer arises from their claim that inanimates can only be good external arguments if they can be construed as agents (like self-propelled cranes) or as inanimate causers (like falling axes, which can affect things in the same way storms can). They claim that an inanimate like a rag can never be either (see their complete set of examples in their 34a-e):
(19)#THIS rag cleaned the dishes. (personally I find this fine)
So, rag should never be a good subject. This is false:
(20) The rag muffled the sound.
(21) The rag smothered the fire.
(22) A rag covered the body.
These are clearly external arguments, because they can passivize (the sound was muffled by the rag, the fire was smothered by the rag, the body was covered by a rag). But this should not be possible, if rags cannot be agents or causers. In contrast, the Initiator view predicts these facts without difficulty. Rags can muffle sound, they can smother fires, and they can cover things. That is, their physical properties are such that they can initiate muffling, smothering, and covering events.
In sum, none of Alexiadou and Schafer's arguments for distinguishing agents from causers goes through. More broadly, I have seen no good reason to divide external arguments into different categories. The "underspecification" theory, which I have called the Initiator view here, accounts for all of the facts straightforwardly.
Alexiadou, Artemis and Florian Schafer (2006), Instrument Subjects Are Agents or Causers. In Proceedings of the 25th West Coast Conference on Formal Linguistics. Somerville, MA: Cascadilla Proceedings Project: 40-48.
Nilsen, Don L.F. (1973), The Instrumental Case in English: Syntactic and Semantic Considerations. The Hague: Mouton.
Ramchand, Gillian Catriona (2008), Verb Meaning and the Lexicon: A First-Phase Syntax. Cambridge: Cambridge University Press.