Ecological food webs define the feeding patterns of interacting species. The architecture of such networks may be affected by dynamical processes operating within them, ultimately influencing the capacity of the networks to persist. As yet relatively little is known about these effects. We compared the architecture of ecological networks with a fixed number of species, constructed in four contrasting ways: (I) topological networks, which required only that species had prey to eat; (II) persistent networks, in which species had also to persist under a simple model of population dynamics; (III) assembled networks, built up by sequential addition of species with dynamical persistence at each step in the sequence; (IV) evolved networks where, in addition to dynamical persistence, body size of species was determined by a simple mutation-selection process. Dynamics had fundamental effects on architecture, the networks of classes II, III and IV being restricted to a small number of trophic levels, in contrast to the non-dynamic, topological class I networks. Class III assembled networks tended to have fewer trophic levels and a more pyramidal biomass distribution than networks of classes II and IV. In evolved class IV networks, the smallest consumers converged to similar body sizes, whereas larger consumers evolved more slowly and did not show such convergence. The results indicate that dynamics affect the architecture of food webs, and that assumptions about simultaneous arrival, sequential arrival and evolution lead to different outcomes. Sequential assembly was shown to have a special property of finding rare sets of persistent species in a small number of steps, suggesting that the rarity of stable communities is not a serious problem in the development of complex communities.