This project studies risk-aware multi-agent reinforcement learning for planning under uncertainty. Standard reinforcement learning optimizes expected return, which can hide rare but severe failures. This is especially problematic in multi-agent systems where agents share resources, induce coupled failure modes, or must coordinate without direct communication. The project will develop and evaluate algorithms for decentralized and centralized learning under recursive entropic risk, CVaR/EVaR-style tail-risk diagnostics, and shared-resource constraints. The first part of the project uses exact finite MDPs (Markov decision processes) and multi-agent MDP baselines. These baselines make it possible to compare learned policies against the true risk-sensitive optimum and to identify when decentralized training is exact, conservative, or misleading. The second part scales the experiments to vectorized multi-agent reinforcement learning environments, including custom shared-risk tasks and PufferLib-based benchmarks. The computational work will consist of controlled sweeps over risk parameters, random seeds, environment variants, and learning algorithms. The requested NAISS resources will be used for GPU-accelerated deep reinforcement learning experiments, CPU-based exact baselines, and reproducible hyperparameter and seed sweeps. The expected outcome is a set of empirically validated methods and benchmark environments for studying how risk-sensitive objectives interact with decentralization, communication assumptions, and shared stochastic failures in multi-agent systems. The project is part of my doctoral studies at Chalmers University of Technology. My main supervisor is Anna Gautier, Chalmers University of Technology.