A protein is identified by a finite sequence of amino acids, each of them chosen from a set of 20 elements. The Protein Structure Prediction Problem is the problem of predicting the 3D native conformation of a protein, when its sequence of amino acids is known. This problem is fundamental for biological and pharmaceutical research. All current math- ematical models of the problem are affected by intrinsic computational limits. In particular, simulation-based tech- niques that handle every chemical interaction between all atoms in the amino acids (and the solvent) are not feasible due to the huge amount of computations involved. These programs are typically written in imperative languages and each approach is based on a particular energy function. There is no common agreement on which is the most re- liable energy function to be used. In this paper we present a novel agent-based framework for ab-initio simulations. Each amino acid of an input pro- tein is viewed as an independent agent that communicates with the others. The framework allows a modular repre- sentation of the problem and it is easily extensible for fur- ther refinements and for different energy functions. Sim- ulations at this level of abstraction allow fast calculation, distributed on each agent. We provide an implementation using the Linda package of SICStus Prolog, to show the feasibility and the power of the method. The code is in- trinsically concurrent and thus natural to be parallelized.