A research-writer-validator crew is naturally both hierarchical and sequential. Consensus or random routing wastes specialization and increases handoff ambiguity. In a GPU-backed agent deployment, the combination of Options A and D maps closest to how the NVIDIA stack expects orchestration, inference, and control policies to be separated. Together, A states “Sequential pipeline coordination with crew-based structured handoffs”; D states “Hierarchical coordination with crew-based task delegation”, so the answer covers both sides of the requirement instead of solving only the model or only the infrastructure layer. The practical pattern is role separation, shared state, structured messages, and explicit handoff contracts between agents. This lines up with NVIDIA guidance because the NVIDIA agent stack is built for composability: agents, tools, and workflows can be profiled and optimized as reusable components. The distractors fail because a fixed pipeline cannot adapt when new evidence arrives, while a monolithic agent makes root-cause analysis painful. This is exactly where NVIDIA’s stack is strongest: separating acceleration, orchestration, policy, and observability.