A crucial activity in project management, software effort estimation (SEE) affects risk assessment, resource management, and budget allocation. The usefulness of hybrid machine learning approaches in improving the precision of effort estimation for agile software development projects is assessed in this study. We suggest three methods: (1) an optimized ensemble using Genetic Algorithms (GAs); (2) an ensemble integrating Bidirectional LSTM with Random Forest, XGBoost, and Voting Regressor; and (3) a Bidirectional Long Short-Term Memory (LSTM) model. Three agile datasets are used to assess the suggested approaches. The Bidirectional LSTM efficiently captures sequential dependencies; however, it performs better when combined with machine learning models, according to performance metrics like MMRE, MdMRE, and Pred (25). By choosing the best model combinations, further GA optimization increases accuracy. Our results highlight the promise of hybrid strategies that combine optimization, ensemble learning, and deep learning. These techniques demonstrate scalability and robustness, attaining excellent accuracy on various datasets and proving their feasibility for real-world software effort estimation applications.