There is scarce evidence on whether and how assistance in humanitarian emergencies and conflict settings impacts household well-being and behaviour. Conducting rigorous impact evaluations in such settings poses multiple challenges in design and data collection. In SEEDS, we evaluate the impact of a complex large-scale multi-arm agricultural intervention on productivity, food security, and resilience in the context of an on-going humanitarian crisis in Syria. Specifically, we identify the causal impacts of agricultural asset transfers over various time horizons (the short-, medium-, and long-run), and across different conditions and subgroups (gender and conflict intensity) at the household-level. We evaluate the effectiveness of irrigation rehabilitation separately at the community-level. We use and combine various data sources, including a unique survey panel dataset collected over a period of four years from multiple governorates in Syria, satellite remote-sensing data, and publicly available violent conflict incidence and weather data. Our findings from using cutting-edge machine and deep learning approaches together with innovative balancing and analytical methods can be summarised as follows: For average treatment effects at the household-level, we find that the provision of agricultural asset support leads to significant improvements in food security in the short- and long-term, three years after the intervention. The positive and significant effect on food security is driven mainly by the increased consumption of healthy food items such as vegetables. In the long-run, livestock support reduces the use of harmful coping strategies households employ to deal with food shortages. Interestingly, we find that households who received vegetable kits are not just less likely to sell their productive assets in the long-term but also are less likely to marry off their young daughters or send their children to work. Overall, we find that both agricultural and livestock asset support is key to improving households’ resilience in the long-term. The irrigation rehabilitation interventions at the community-level positively affected agricultural productivity compared to the pre-intervention and pre-conflict periods. However, these effects were only significantly pronounced in the spring season. As for the heterogeneity analysis, we find that female-headed households benefit remarkably more in terms of food security in the medium-term compared to male-headed families. Moreover, households residing in areas that are moderately affected by violent conflict show stronger food security improvements compared to households from peaceful or conflict-intense settings. Overall, we draw three overarching lessons from our findings in SEEDS: First, agricultural support in protracted conflict settings effectively improves the long-term welfare and resilience of vulnerable households. In fact, the presence of an ongoing humanitarian operation acts as a social safety net if circumstances deteriorate suddenly. Second, not all interventions are equally effective, and not all households equally benefit, underscoring the need to design and implement inclusive context-specific interventions with detailed targeting. Third, methodologically, using multiple remote data sources and machine learning methods help overcome challenges in conducting rigorous impact evaluations in hard-to-reach humanitarian emergency settings.