When Agents Fail: Why Multi-Criteria Decision Frameworks Outperform Autonomous AI Agents for High-Stakes Decisions

The rapid deployment of autonomous AI agents in production environments has exposed critical vulnerabilities that challenge fundamental assumptions about agent reliability. Shapira et al. (2026) documented 10 security vulnerabilities and 6 emergent safety behaviors across 6 autonomous agents observed over 14 days in unstructured multi-agent environments (arXiv:2602.20021). These failures -- including disproportionate response, manipulation via social engineering, identity hijacking, infinite resource loops, and constitutional corruption -- are not implementation bugs but architectural consequences of delegating decisions to stochastic generative models. We perform a systematic mapping of all 16 documented case studies to the AEGIS multi-criteria decision framework, demonstrating that every vulnerability class maps to a structural guarantee in MCDA: proportionality via normalized TOPSIS distances, manipulation resistance via fixed criteria schemas, deterministic reproducibility, bounded resource consumption, and complete audit trails. We argue that for domains requiring auditability, proportionality, and adversarial robustness -- including cybersecurity, finance, healthcare, and critical infrastructure -- structured MCDA frameworks like AEGIS are architecturally superior to autonomous agents for the decision layer, while AI remains valuable for feature extraction and enrichment.