Context,Scaling agricultural innovations is increasingly recognised as a political rather than merely technical process, yet power dynamics remain systematically under-theorised in agricultural research for development (AR4D). Existing frameworks rarely account for how power shapes agenda-setting, knowledge validation, benefit distribution, and actor authority across the scaling cycle.,Objective,This paper addresses the existing gap by developing and applying an integrated analytical framework combining three complementary theoretical approaches: the Multi-Level Perspective (MLP), Gaventa’s Power Cube, and Avelino and Wittmayer’s power typology to analyse the operation of power across the agricultural scaling cycle.,Methods,Drawing on a narrative literature review and a literature-informed AI-assisted case study of the fictional Kijani Valley Climate-Smart Initiative in the Global South, we map actor groups across niche, regime, and landscape levels of the MLP and assess how expressions of power (power over, to, with, and within) function across closed, invited, and claimed spaces at each level.,Results and conclusions,Our findings reveal that scaling outcomes emerge from dynamic negotiations of power across levels. Landscape-level actors, particularly donors and global policy discourses, exert largely hidden and invisible power through funding architectures and performance metrics. Regime actors, including ministries and private sector actors steer diffusion through formal authority and informal coalition building. Niche level actors such as women's groups, youth enterprises, and farmer–researcher learning platforms claim spaces, build collective agency, and strengthen internal capacities. We find that scaling effectiveness and equity are co-produced outcomes of these multi-level power negotiations rather than separate objectives.,Significance,Power-aware scaling is not an ethical add-on but a functional requirement for sustained agricultural transformation and progress toward the Sustainable Development Goals. This paper offers practical entry points for integrating power analysis into AR4D programme design and calls for longitudinal actor-mapping and revised monitoring frameworks that capture relational and collective dimensions of power alongside conventional reach and adoption metrics.