Capability-Fenced Orchestration of LLM Multi-Agent Systems: An Empirical Evaluation of Reliability, Safety, and Cost Trade-offs

Abstract

Large language model (LLM) agents that plan, call external tools, and coordinate with one another are increasingly composed into orchestrated multi-agent systems (MAS). While prior work argues qualitatively that orchestration improves reliability and that capability scoping mitigates agentic risk, controlled measurements that isolate the contribution of individual orchestration controls remain scarce. This paper presents an empirical study of OrchGuard, a capability-fenced orchestration framework that combines least-privilege capability tokens, inter-stage validation gates, and a verify-before-commit policy engine over a six-layer reference architecture. We evaluate OrchGuard against a single-agent baseline and an unguarded multi-agent baseline on two task batteries, enterprise procurement orchestration and scientific literature synthesis, across 1,400 benign and 1,200 adversarial trials, and we report a component ablation. Relative to the single-agent baseline, OrchGuard raised mean task success from 53.5% to 92.0% and reduced output-quality variance by roughly an order of magnitude. Under an indirect-prompt-injection suite, OrchGuard reduced attack success from 71% to 11% and excessive-agency violations from 23 to 2 per 100 tasks, at a cost overhead of about 2.7x and a latency overhead of about 3.7x over the single-agent baseline. The ablation indicates that capability fencing contributes most to attack resistance while validation gates contribute most to task success, and that the two controls are largely complementary. We conclude that the safety and reliability benefits of orchestration are attributable to specific, separable runtime controls rather than to multi-agent structure alone, and we discuss the resulting reliability-versus-utility and cost trade-offs.