Modern enterprises rely on real-time streaming pipelines for continuous data processing. Traditionally, these pipelines have static logic, but generative AI agents enable adaptive execution. This allows workflows to dynamically construct plans and take actions based on data content. For instance, an angry customer message can trigger an order lookup, remediation, and customer notification. However, running generative AI workflows at scale faces challenges like high latency and cost. The proposed solution combines Google Dataflow with the Agent Development Kit (ADK) to create a hybrid pipeline, efficiently handling high-volume streams. A lightweight upstream model filters events, routing only complex cases to a downstream generative AI agent. This pre-filter plus agentic action pattern is universally applicable to streams with mostly routine events. Examples include IT operations log triaging, financial fraud detection, and industrial IoT monitoring. The architecture leverages the cost and speed of local CPU models with the advanced capabilities of generative AI agents. By employing a CPU-bound model for initial filtering, only a small fraction of events incur the cost of using heavyweight generative AI models. This hybrid approach significantly reduces costs and maintains high streaming throughput. The integration with Apache Beam and the ADK ensures maintainable and clean code.