Local traffic can be offloaded through device-to-device (D2D) communication, but mode, spectrum, power and propagation-path decisions remain coupled when cellular protection, finite queues and reconfigurable intelligent surface (RIS) assistance are considered. An explainable distributed artificial intelligence framework is proposed in which each D2D pair is represented by a Belief-Desire-Intention agent with extended capabilities (BDIx). Beliefs are formed from channel, queue, interference, service and RIS-quality observations, while desires are formed from rate, reliability, delay, energy and fairness requirements. Direct D2D, RIS-assisted D2D, cellular fallback and defer plans are ranked locally, followed by common feasibility repair, minimum-power adaptation and executed-action explanation, in which the final executed action is associated with its dominant weighted unmet desire. The policy is strengthened with RIS-aware plan rewards and an adaptive conservative belief margin for imperfect channel state information (CSI); physical channel equations remain unchanged. Candidate configurations are selected on disjoint development and validation seeds before the selected policy is frozen. On ten internal-control seeds, mean throughput of 28.86 Mbit/s, a packet delivery ratio (PDR) of 99.27%, total latency of 20.82 ms and terminal-plus-RIS energy efficiency of 14.53 Mbit/J are obtained by BDIx-RIS. On ten independent published-approach comparison seeds, the corresponding means are 28.92 Mbit/s, 99.57%, 19.06 ms and 13.44 Mbit/J. Nine numerical comparator structures are evaluated: eight paper-specific optimizers and one explicitly identified adapted multi-agent deep Q-network (DQN) decision-structure comparator. Relative to swap matching with successive convex approximation (SCA), latency is reduced by 15.28% and terminal-plus-RIS energy efficiency is increased by 106.81%; the Holm-corrected paired Student t and Wilcoxon p-values are 0.00512 and 0.01758 for latency and 3.24 × 10^-5 and 0.01758 for terminal-plus-RIS energy efficiency, respectively, while throughput and PDR differences are directionally positive but are not statistically resolved after correction. Statistically supported gains are therefore established for latency and terminal-plus-RIS energy efficiency, while positive throughput and PDR differences are reported as numerical trends. A referenced 2020 to 2026 literature comparison, CSI robustness, RIS operating-region analysis, ablations and explanation-faithfulness tests are used to establish the operating conditions under which the gains persist and to show that RIS benefits remain geometry and load dependent under the declared common model.

BDIx agents, device-to-device communication, distributed artificial intelligence, explainable decision making, reconfigurable intelligent surface, resource allocation, 5G, 6G