We document how H-2A job orders describe task and skill content for U.S. seasonal agriculture from FY2020 to FY2026, using 128,599 job orders and 804,296 farm-like sentences across the three Standard Occupational Classification (SOC) codes that hold 96.8 percent of certified worker positions. Token-level analysis finds rising equipment and maintenance vocabulary alongside falling hand-tool vocabulary. A supervised labelling onto the 52 Occupational Information Network (O*NET) ability and task definitions finds the FY2026 level distribution agrees with O*NET’s physical and psychomotor core, while the fastest movers are cognitive and sensory abilities O*NET ranks as low-importance. Unsupervised clustering surfaces residue task clusters absent from O*NET, including mechanized orchard platforms and trellis-and-vine training. We read the disagreement as evidence that O*NET profiles for these occupations lag behind what they actually involve. Workforce training would benefit from adding mechanized platform operation, trained-canopy maintenance, and English-language instruction-following capabilities.
