Agricultural misinformation on social media poses risks to Indonesian food security and farmer livelihoods. False claims about fertilizers, pesticides, crop varieties, and farming practices can spread rapidly across social media platforms such as Instagram, YouTube, and Facebook, as well as through publicly available datasets, potentially influencing agricultural decisions and outcomes. This paper introduces AgriMisinfo-ID, a bilingual, multi-platform dataset for agricultural misinformation detection, containing 5,288 labeled samples collected from social media and supplementary public datasets across Indonesian agricultural contexts. A hybrid detection system is proposed that combines an ensemble of fine-tuned Transformer models, IndoBERT and XLM-RoBERTa, with a Knowledge Verification module that cross-references agricultural claims against Wikidata and Wikipedia. Training uses Focal Loss to address class imbalance, together with GPT-4o-mini-based paraphrase augmentation for minority classes. Across three random seeds, the weighted ensemble achieves an F1-Macro of 0.5870 ± 0.0028 and an accuracy of 0.8027 ± 0.0039 on the test set, outperforming the individual models, a TF-IDF/SVM baseline, and an equal-weight ensemble in terms of F1-Macro. The Knowledge Verification module provides evidence-based verdicts that can support the inspection and auditability of model decisions. This work provides a reproducible benchmark for agricultural misinformation research in bilingual, low-resource settings.