Five experiments, using a subject-verb agreement error elicitation procedure, investigated syntactic planning processes in production. The experiments examined the influence of semantic integration--the degree to which phrases are tightly linked at the conceptual level--and contrasted two accounts of planning: serial stack-based systems and parallel activation-based systems. Serial stack-based systems rely on memory-shifting processes to coordinate ongoing planning. Memory-shifting should be easier for more integrated phrases, resulting in fewer errors. Parallel, activation-based systems, on the other hand, maintain multiple representations simultaneously in memory. More integrated phrases will be more likely to be processed together, resulting in increased interference and more errors. Participants completed stimuli like "The drawing of/with the flower(s)," which varied local noun number ("flower(s)") and the relationship between the head (drawing) and local noun. In some constructions, the nouns were tightly integrated (e.g., "of"), whereas in others the relationship was looser (e.g., "with," specifying accompaniment). In addition to the well-established local noun mismatch effect (more errors for plural than for singular local nouns), all experiments revealed larger mismatch error effects following tightly integrated stimuli. These results are compatible with parallel activation-based accounts and cannot be explained by serial, memory-shift-based accounts. The experiments and three meta-analyses also ruled out alternative accounts based on plausibility, argumenthood, conceptual number, clause packaging, or hierarchical feature-passing, reinforcing the general finding that error rates increase with degree of semantic integration.