Traditional electricity grid networks have approached critical resilience thresholds, with archaic grid architectures demonstrating inadequate capacity to accommodate volatile demand fluctuations. South Africa’s current energy crisis, characterised by unpredictable power outages and extensive economic losses from electricity outages, emphasises the need for novel optimisation frameworks. This research proposes a framework named AMPERE (Autonomous Multi-Agent Predictive Electricity Reinforcement Learning Engine), a dual-phase solution which combines hierarchical multi-agent deep reinforcement learning (MADRL) and evolutionary algorithms (EAs) in order to address electricity distribution challenges through decentralised multi-agent coordination. A significant gap exists in research integrating distributed hierarchical knowledge-sharing MADRL with EAs at both regional and national levels, since current approaches focus on singular artificial intelligence (AI) methodologies or non-hierarchical frameworks. Ultimately, AMPERE introduces a novel three-tier hierarchical agent architecture consisting of smart grid national agents, smart home regional agents, and smart home battery agents, utilising Pecan Street’s real-world Internet of Things (IoT) time-series datasets. The Centralised Training with Decentralised Execution (CTDE) paradigm enables effective multi-agent coordination across regional and national scales, subsequently enhancing smart grid stability and scalability while minimising electricity lost in distribution.