Abstract
<title>Abstract</title> <p>Cloud–edge AI agents depend on model, retrieval, policy, and tool services whose partial failure can turn ordinary availability mechanisms into unsafe continuation. We introduce DMO-AI, a risk-aware degraded-mode orchestrator that selects normal, rate-limited, cached, local, read-only, handoff, or blocked execution from task risk and observed dependency health. We first evaluate DMO-AI in a single-host containerized testbed with real HTTP chains and Toxiproxy faults: two frozen stages cover 141,000 executions, conventional baselines, ablations, sensing, and overhead. We then perform a complete 8,000-event validation across an x86-64 cloud node and ARM64 edge node running K3s, Istio/Envoy, real SmolLM2 inference, and a transactional SQLite tool. In the multi-node study, DMO-AI with Istio achieved 93.95% safe completion versus 64.70% for Istio alone, a paired 29.25-percentage-point gain (95% bootstrap CI 27.30–31.85), with no observed policy violations, duplicate commits, negative inventory, or high-risk edge substitutions. Mean latency increased from 139.67 to 158.83 ms. The results support explicit risk-bounded degradation as a semantic control layer that complements transport resilience without claiming production-wide generality.</p>