Abstract
<jats:p>Recent advances in multimodal generative models have enabled high-quality synthesis across images, videos, audio, 3D content, and document-centric media. Yet most systems still treat generation as a direct prompt-to-artifact mapping or a fixed pipeline, limiting their ability to handle long-horizon objectives, compositional constraints, tool coordination, and iterative revision. At the same time, agent-based systems show that complex tasks can be addressed through explicit goals, planning, tool use, memory, and feedback-conditioned decisions. This convergence suggests that the central challenge is no longer only how to improve individual generators, but also how to control the generation trajectory that connects intent, intermediate artifacts, and final outputs. In this survey, we formalize this perspective as Agentic Generative Systems (AGS): generation-centered systems that represent content creation goals, decompose them into intermediate steps or structured representations, and orchestrate generative models, tools, or executable operations along a multi-step trajectory. Depending on capability level, an AGS may execute an open-loop plan, revise outputs through local feedback, or strategically replan using persistent state and memory. We introduce a three-level capability hierarchy and a mechanism-centered taxonomy covering agent architectures, planning and control, tool orchestration, memory, and feedback. Under this framework, we review representative systems across image, video, audio, 3D, and document-centric generation, showing that the same agentic progression recurs across modalities while the object of control changes from spatial composition to temporal coherence, event timing, world state, or information structure. We further consolidate datasets, benchmarks, and evaluation protocols for AGS, emphasizing both final artifact quality and trajectory-level behavior. Finally, we discuss open challenges in long-horizon consistency, process-level evaluation, reliability, safety, cost-aware control, and generalization. This survey provides a unified conceptual foundation for studying autonomous multimodal content creation as a system-level decision process. We maintain a curated repository at https://github.com/xxlbigbrother/Awesome-Agentic-Generative-Systems.</jats:p>