The antagonistic pleiotropy theory of senescence postulates genes or traits that have opposite effects on early-life and late-life performances. Because selection is generally weaker late in life, genes or traits that improve early-life performance but impair late-life performance should come to predominate. Variation in the strength of age-specific selection should then generate adaptive variation in senescence. We demonstrate this mechanism by comparing early and late breeders within a population of semelparous capital-breeding sockeye salmon (Oncorhynchus nerka). We show that early breeders (but not late breeders) are under strong selection for a long reproductive lifespan (RLS), which facilitates defence of their nests against disturbance by later females. Accordingly, early females invest less energy in egg production while reserving more for nest defence. Variation along this reproductive trade-off causes delayed or slower senescence in early females (average RLS of 26 days) than in late females (reproductive lifespan of 12 days). We use microsatellites to confirm that gene flow is sufficiently limited between early and late breeders to allow adaptive divergence in response to selection. Because reproductive trade-offs should be almost universal and selection acting on them should typically vary in time and space, the mechanism described herein may explain much of the natural variation in senescence.