Digital Twins (DTs) provide a networked, data-enabled approach to safer control in Controlled-Environment Agriculture (CEA), but definitions along with stages of maturity and deployable patterns remain ambiguous. This paper characterizes DT variants specific to CEA: digital model, digital shadow, and bi-directional twins. We provide a reference architecture that includes sensing/ingestion, a resilient data backbone, hybrid (biophysical + Machine Learning (ML)) modeling, a scenario/what-if layer, and safe actuation with a human in the loop. We highlight recent applications in CEA (climate and fertigation control—fertilizer-in-water dosing), summarize reported benefits (improved resource efficiency and uniformity) and reported gaps (interoperability, calibration, external validity). We outline an ‘intelligence-in-the-loop’ where process models coupled with reinforcement learning, produce risk-aware recommendations. Finally, we provide a staged Minimum Viable CEA DT roadmap and associated key performance indicators (KPIs) for water, energy, and food quality. In summary, our findings clarify DT and architecture for CEA while also acknowledging limitations of sensor coverage, latency budgets and audit requirements for calibration and governance.