Our attempts to understand the nature of dark energy are complicated by the fact that, beyond the cosmological constant, we lack a single compelling model for how it may evolve. Here we develop a flexible and efficient non-parametric Bayesian method for reconstructing the time evolution of the dark energy equation of state w(z) from observational data. Of particular importance is the choice of prior, which must be chosen carefully to minimise variance and bias in the reconstruction. We apply it to a collection of cosmological data, and find that the cosmological constant appears consistent with current data, but that a dynamical dark energy model which evolves from w<~1 at z~0.25 to w > -1 at higher redshift is mildly favored. Estimates of the Bayesian evidences show little preference between the cosmological constant model and the dynamical model for a range of correlated prior choices. Looking towards future data, we find that the best fit models for current data could be well distinguished from the ΛCDM model by observations such as Planck and Euclid-like surveys.