Advances in large language models (LLMs) and generative AI have enabled software agents that can plan, call external tools, and coordinate with one another to solve tasks that previously required substantial hand-written control logic. This review traces the transition from developer-driven, manually coded automation toward AI-augmented autonomous orchestration, in which a coordination layer composes collections of specialized agents into cooperative systems. We situate this shift historically, propose a working taxonomy, and describe the technical building blocks of orchestrated multi-agent systems (MAS): the agent, communication, planning, state, policy, and observability layers. We then analyze the expanded attack surface and failure modes these systems introduce, cascading delegation errors, indirect prompt injection, excessive agency, and accountability diffusion, and survey the governance, verification, and evaluation practices that have emerged in response, including capability-scoped access, runtime policy enforcement, immutable provenance logs, and standards such as the NIST AI Risk Management Framework and the EU AI Act. Finally, we present design patterns, a set of evaluation metrics, three illustrative case studies, and a research roadmap. Our central argument is that orchestration is not merely an engineering convenience but a new systems abstraction that relocates complexity from application code into governance, verification, and runtime control, and that responsible deployment therefore depends on a defense-in-depth combination of technical and organizational controls.
Read the full scope of JBSS, including biological sciences, biotechnology, natural products, nanobiotechnology, environmental biotechnology, and sustainability sciences.
Read the policy for correcting the scholarly record, issuing retractions, handling withdrawal requests, expressions of concern, and post-publication updates.