Agglomeration is the process of combining smaller tasks with larger ones in order to improve performance. If the previous two stages of the design process partitioned the problem into a number of tasks that greatly exceed the number of processors available, and if the computer is not specifically designed to handle a huge number of small tasks (some architectures, such as GPUs, handle this fine and indeed benefit from running millions, or even billions, of tasks), then the design can turn out to be highly inefficient.

Commonly, this is because tasks have to be communicated to the processor or thread so that they compute the said task. Most communications have costs that are disproportionate to the amount of data transferred, but also incur a fixed cost for every communication operation (such as the latency, which is inherent in setting up a TCP connection). If the tasks are too small, then this fixed cost can easily make the design inefficient.