We consider the problem of programming parallel applications for a Grid environment, in the presence of the two main challenges (i) high-latency communications and (ii) heterogeneity. We describe a new scripting language, GEL, whose semantics have been designed for execution on a heterogeneous, distributed computer. The language provides syntactic constructs for while loops, conditionals and explicitly parallel execution. The language is designed to work well given these two challenges, and to allow succinct representation of parallel programs, resulting in easier-to-maintain code. The programs can use legacy applications without re-engineering, and do not explicitly refer to resource names or use middleware-specific references. This middleware-independence allows us to execute the same script on an SMP machine, cluster or Grid. We describe three example applications written in GEL: an optimisation problem solved using a swarm algorithm; an allergenicity prediction pipeline; and transcript analysis for tissue-specific gene expression. We have run these scripts unchanged on an SMP machine, on PBS, SGE and LSF clusters, and on a Globus-based Grid.