Mental health issues are among the most neglected problems in public health and globally, and treatment gaps are significant even among the wealthy nations. The development of artificial intelligence and multi-agent systems can change the accessibility and enable privacy-preserving personalization of mental health support. This paper describes a Hybrid Agentic AI Framework which consists of a Random Forest ensemble classifier, DistilRoBERTa-based Emotion Detection Agent, a Severity Stratification Agent, and a FLAN-T5 Language Generation Agent, and combines them in a single, Streamlit-based application. The framework is tested on the OSMI Mental Health in Tech Survey and a public General Mental Health Survey, which allows evaluation of the framework in terms of robustness and generalizability across different datasets. To combat class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) is used and validation set threshold tuning is used to improve the sensitivity of the F1-Score.Experimental results show that the hybrid model achieves accuracy of 93.25% and 97.18% ROC-AUC for the OSMI dataset, and 90.00% accuracy and 96.00% ROC-AUC for the General dataset. It outperforms Logistic Regression, SVM, and Random Forest in all scenarios. Gini-based and permutation feature importance methods show that, for both datasets, the most important features are work interference and family history. The model also achieves zero false negatives in the OSMI test set due to the addition of affective signals based on the Emotion Detection Agent. This work shows that hybrid agentic AI combining structured ML with Conversational Language Generation can yield clinically accurate risk prediction and tailored guidance to the end user, all in one system and ready for deployment.

Keywords: mental health prediction, random forest, agentic AI, FLAN-T5, DistilRoBERTa, SMOTE, threshold optimisation, emotion detection, cross-dataset evaluation, explainable AI, large language models, healthcare AI