<p>Previous work has shown that groups of underutilized IoT devices can be organized into temporary clusters to execute decomposable services collaboratively, using spare CPU, memory, battery, and network capacity without compromising their primary functions. However, earlier orchestration-based approaches rely on an initial allocation of work proportional to each device’s estimated capacity. Such static or semi-static allocation becomes inefficient when task costs are irregular, nodes are heterogeneous, background load varies, or devices join and leave during execution. This paper proposes a QoS-aware distributed work-stealing model for opportunistic IoT clusters. The proposed approach replaces centralized work assignment with local double-ended queues, capacity-aware victim selection, lease-based task ownership, bounded stealing, and QoS-preserving admission control. We formalize the system model, define residual capacity and utility-aware stealing decisions, provide the core distributed algorithm, and analyze termination, safety, bounded interference, and expected makespan under standard work-stealing assumptions extended with communication and heterogeneity costs. We evaluate the proposed approach against five baselines—proportional static decomposition, multi-round proportional decomposition, centralized dynamic rescheduling, random stealing, and capacity-aware stealing (a decentralised scheduler with the same capacity-aware victim selection but without QoS admission, included to isolate the contribution of QoS-aware control)—across heterogeneous IoT cluster sizes ranging from 4 to 64 devices, three task-cost distributions (uniform, log-normal, Pareto), and three failure regimes (none, transient, permanent). The results demonstrate that QoS-aware distributed work stealing matches the makespan of the evaluated centralised dynamic scheduler without placing it on the critical path, scales without performance degradation across the evaluated range of up to 64 devices, where uncoordinated random stealing breaks down completely; achieves the highest and most consistent cluster utilisation; and degrades gracefully under failures, sustaining task completion where the centralised and proportional baselines leave the most work unfinished, even under an aggressive stress regime where approximately 90% of devices fail during a typical run. These benefits come at the cost of a higher control-plane overhead from periodic status dissemination; we analyse this cost explicitly and identify gossip-based dissemination as the means to extend the approach beyond the evaluated scale. These results confirm that distributed work stealing preserves the central goal of the previous framework—exploiting otherwise unused IoT capacity to improve edge service execution within acceptable QoS ranges—while improving load balance, resilience, and scalability for irregular workloads.</p>