Seedbed requirements for the establishment of grasses and legumes were studied in 3 experiments conducted over 3 years (1981-83) on the sandy ( 20 mg kg-1 Bray P) ultisol soils found in the dry tropical rangelands of western Hainan Island, China. Density at the end of the wet season and production accumulated over the growing season (May-October) were the discriminative parameters used. Experiment 1 measured the effect of seedbed preparation (cultivated and uncultivated but heavily grass pasture) on the establishment of 2 grasses, buffel (Cenchrus ciliaris L. cv. Biloela) and sabi (Urochloa mosambicensis (Hack.) Dandy cv. Nixon), sown alone or mixed with 3 stylos (Stylosanthes guianensis (Aubl.) Sw. var. guianensis cv. Graham, S. hamata (L.) Taub. cv. Verano and S. scabra Vog. cv. Seca) on the 2 soil types. Biloela buffel established better and produced more in subsequent years than Nixon sabi grass when sown on a cultivated seedbed, but neither grass established when sown into undisturbed grassland. Companion stylos established on both seedbeds, but initial density and yield was lower on the uncultivated seedbed. However, in 3-year-old swards, stylo yield on uncultivated seedbeds equalled or exceeded yield on cultivated seedbeds. Seca was the most productive stylo, but Graham yielded more consistently between years. Soil type had no effect on sown grass production, but legumes grew better on the sandy soil. Experiment 2 measured the effect of 4 seedbeds (heavily grazed, 1 disking, sprayed with herbicide, and cultivated to fine tilth) on establishment on Verano and Seca stylos on sandy soil. Density of both stylos was highest where grass competition was minimized with herbicide and lowest where heavy grazing was imposed prior to seeding. Compared to herbicide, disking resulted in lower density and yield in the establishment year, but the difference disappeared with time. Burial of seed too deeply and subsequent competition from native species account for the poor performance of both stylos when drilled into a cultivated seedbed which was still evident in 3-year-old swards. Experiment 3 examined the interaction between seedbed type (heavily grazed, burnt, 1 disking, sprayed with herbicide) and superphosphate application. Superior establishment was measured on herbicide and disked treatments and production was still higher in 3-year-old swards than heavy grazing. Burning was more effective than heavy grazing. Increasing fertilizer input improved legume production on all seedbeds, but did not affect density. Reduction in legume yield between 2- and 3-year-old swards in the burnt and disked treatment due to grass competition highlights the need for appropriate grazing to maintain legume content of augmented grasslands. Overall we recommend rough disking as the most practical and cost-effective means to augment native grasslands with stylos. A fine seedbed is recommended to provide the disturbance required to establish grasses. There is no advantage gained by including grasses with legumes oversown into living grassland.