Modern power grids face a multitude of challenges: A high share of renewables means a more sophisticated management of real power demand and supply, ancillary services are provided in an ever more decentralized manner, and power grids must become resilient instead of just being robust. Agent systems have established themselves as methodology for a decentralized and resilient operation of modern power grids. Especially learning agents based on Deep Reinforcement Learnings can react to unforseen events and find good strategies even in complex situations. In this lecture, we will introduce an approach for flexibility modelling as a way to provide an agent's view of the world, and will extensively concern ourselves with the application of Deep Reinforcement Learning in power grids, including approaches to explainability and learning from domain knowledge (offline learning).