Interaction nets are a low level graphical computation paradigm first proposed by Yves Lafont and based on Jean-Yves Girard's proof nets for linear logic. An interaction net system comprises: a set of agents, each with one principal port and zero or more auxiliary ports; a set of rules between agents (there is at most one rule for every pair of agents); and a net on which the rules are to be applied. Compared to traditional term syntax and linear graph grammars, interaction nets enforce linearity -- each resource is used exactly once --, from which we can derive strong confluence. Thus, they provide a natural language for massive parallelism.
They are also at the heart of the efficient and optimal, in Levy's sense, evaluators for lambda calculus available today.