Plant diversification through crop rotation or agroforestry is a promising way to improve sustainability of agroecosystems. Nonetheless, criteria to select the most suitable plant communities for agroecosystems diversification facing contrasting environmental constraints need to be refined. Here, we compared the impacts of 24 different plant communities on soil fertility across six tropical agroecosystems: either on highly weathered Ferralsols, with strong P limitation, or on partially weathered soils derived from volcanic material, with major N limitation. In each agroecosystem, we tested several plant communities for diversification, as compared to a matching low diversity management for their cropping system. Plant residue restitution, N, P and lignin contents were measured for each plant community. In parallel, the soil under each community was analyzed for organic C and N, inorganic N, Olsen P, soil pH and nematode community composition. Soil potential fertility was assessed with plant bioassays under greenhouse controlled climatic conditions. Overall, plant diversification had a positive effect on soil fertility across all sites, with contrasting effects depending on soil type and legumes presence in the community. Communities with legumes improved soil fertility indicators of volcanic soils, which was demonstrated through significantly higher plant biomass production in the bioassays (+18%) and soil inorganic N (+26%) compared to the low diversity management. Contrastingly, communities without legumes were the most beneficial in Ferralsols, with increases in plant biomass production in the bioassays (+39%), soil Olsen P (+46%), soil C (+26%), and pH (+5%). Piecewise structural equation models with Shipley's test revealed that plant diversification impacts on volcanic soil fertility were related to soil N availability, driven by litter N. Meanwhile, Ferralsols fertility was related to soil P availability, driven by litter P. These findings underline the importance of multifactorial and multi-sites experiments to inform trait-based frameworks used in designing optimal plant diversification in agroecological systems.