We describe the development of a new object kinetic Monte Carlo code where the elementary defect objects are off-lattice atomistic configurations. Atomic-level transitions are used to transform and translate objects, to split objects and to merge them together. This gradually constructs a database of atomic configurations- a set of relevant defect objects and their possible events generated on-the-fly. Elastic interactions are handled within objects with empirical potentials at short distances, and between spatially distinct objects using the dipole tensor formalism. The model is shown to evolve mobile interstitial clusters in tungsten faster than an equivalent molecular dynamics simulation, even at elevated temperatures. We apply the model to the evolution of complex defects generated using molecular dynamics simulations of primary radiation damage in tungsten. We show that we can evolve defect structures formed in cascade simulations to experimentally observable timescales of seconds while retaining atomistic detail. We conclude that the first few nanoseconds of simulation would be better performed using molecular dynamics, and that, in the 20keV PKA cascades annealing simulations we consider here, the first few microseconds establish the conditions for long timescale evolution.