This paper reports a rather striking distinction found in the performance of agrammatic patients: Their comprehension deficit distinguishes not only different syntactic constructions, but also verbs with different thematic structure. Thus, the coupling of a variety of sentence types with different verb types yields surprising performance patterns, demonstrating once again that the deficit these patients suffer from is highly selective from a grammatical viewpoint. This pattern of selectivity has important implications for linguistic theory, for it describes an error pattern that can be accounted for only by a theory that assumes a (normal) lexicon that encodes thematic labels and refers to a thematic hierarchy. In addition, the results argue for the necessity of a cognitive strategy (first NP = Agent) as part of the structural account of agrammatic comprehension (the Trace-Deletion Hypothesis). It is shown that, contrary to some recent claims, an account lacking such a strategy cannot derive agrammatic performance patterns properly. New data are presented, from an experiment that used an anagram task to test agrammatic comprehension of active and passive sentences containing verbs of three different thematic types. Verbs varied in that their external argument was Agent, Instrument, or Experiencer. The finding was that on all active sentences, performance was above chance, whereas on the passives, performance split: on verbs with Agent or Instrument external arguments, performance was at chance, corroborating previous findings. On Experiencer verbs, though, performance was below chance. These data argue decisively for the TDH as originally presented, namely, for an account that assumes trace deletion, coupled with a strategy. The implications of these results to the theory of brain/language relations are discussed, in the context of the theory of lexical representation.