The use of synthetic data for training machine learning models (ML) in social media domains can address issues such as data availability and bias, but poses challenges, including properly reflecting causal relationships and matching the consistency of real data. In this paper, we explore the benefits and limitations of using synthetic data generated by cognitive agent simulations. By simulating human interactions and social media dynamics, these models can capture constraints and nuances of real-world scenarios. We report initial experiments that show that ML algorithms trained on real data augmented with synthetic data outperform those trained solely on original data, achieving up to 25% improvement in KS distance and RMSE metrics. This approach is applied to two domain problems: predicting code quality based on open-source code discussions and detecting and countering bot attacks on social media platforms. For code quality prediction, we used discussions and patches from the Linux Kernel Mailing List to predict patch reversions. In the bot attack detection problem, synthetic Reddit data helps create realistic social network environments to study interactions between influencers and bots under different conditions. The paper presents empirical evidence supporting the effectiveness of synthetic data in improving ML model performance and introduces an agent-based framework for generating realistic synthetic data for social media experiments. The findings suggest promising avenues for future research and highlight the potential of this approach.