Efficient data pipelines are essential for transforming raw information into structured outputs that support analytics, decision-making, and machine learning systems. Manual creation of these pipelines can be time-consuming and error-prone, while traditional automation tools often lack adaptability for domain-specific needs or evolving data structures. This gap calls for a more intelligent and context-aware approach to code generation. This work presents a task-specific code synthesis framework for data pipelines using transformer-based large language model (LLM) agents. The proposed system accepts natural language descriptions of pipeline requirements and produces executable code that aligns with the intended data flow, transformation logic, and performance goals. By combining domain-specific fine-tuning, contextual embeddings, and adaptive reasoning, the generated code meets both functional accuracy and operational efficiency. The architecture is trained on a curated dataset covering ingestion processes, transformation scripts, schema alignment patterns, and optimization templates. Generated code undergoes automated validation, including syntax verification, schema consistency checks, and execution profiling. This structured verification process improves reliability while reducing development cycles. Performance is assessed using metrics such as functional correctness rate, execution success rate, latency reduction, and resource utilization efficiency. Experimental results demonstrate significant gains in code quality, execution speed, and adaptability compared to rule-based and template-driven methods. These findings indicate that transformer-based LLM agents can provide a scalable, accurate, and adaptive solution for building high-quality data pipelines in dynamic environments.