Orchestration and Verification of Agentic AI Systems: A Survey of Multi-Agent Collaboration and Safety

The transition from isolated Large Language Models (LLMs) to orchestrated Multi-Agent Systems (MAS) enables complex problem-solving but introduces profound safety, security, and verification challenges. This paper presents a comprehensive survey of orchestration architectures and safety mechanisms in agentic AI systems. This study systematically analyzes collaboration paradigms and orchestration topologies, demonstrating how distributed interactions expand the attack surface. A unified threat taxonomy is proposed, distinguishing between intrinsic agent vulnerabilities, such as prompt injection and cognitive drift, and extrinsic, MAS-specific threats, including cross-agent poisoning and emergent shadow behaviors. Furthermore, we evaluate current defense strategies, from capability sandboxing and runtime monitoring to zero-trust architectures, while highlighting the fundamental limitations of traditional formal verification in non-deterministic environments. By examining applications across software engineering, healthcare, cybersecurity, and finance, we identify domain-specific safety patterns and the necessity of human-in-the-loop oversight. Finally, we discuss regulatory challenges and outline future research directions, emphasizing the need for compositional verification, standardized adversarial benchmarks, and interdisciplinary governance to ensure the trustworthy deployment of autonomous systems.