Agentic AI Factories: Constructing Intelligent Workflows for Data-Driven Decision Making

Abstract In recent years, the concept of AI-driven workflow automation has gained prominence as organizations seek to harness natural language inputs to generate complex workflows. However, these systems produce boiler plan templates and basic workflow configurations. Applied scientists and engineers still need to adapt them, add tooling capabilities and evaluate them for reliable inference and decision making. These systems often operate by simply connecting the output of one model or tool to the input of another without transparently indicating whether each module is adequately pre-trained, prompt engineered or requires RAG etc. In contrast, our proposed agentic AI Agent Factory (AAF) is a framework that provides an interactive method of developing fully working AI Agentic workflows. While applied and data scientists can start defining the workflow requirements in natural language mode, AI factory will spin-off an consultative dialogue with the. AAF starts with providing a high level workflow with potential models and tools. Further, AAF asks clarifying questions to collect additional context such as historical data, quality of inferences, evaluation benchmarking of decisions. This process also ensures that any human-in-loop requirements, as needed ensure quality of inference and decision making. By iteratively constructing workflows and leveraging tools like Agents orchestration, RAG, tools, evaluation tools and human-in the loop interfaces the framework continuously refines its performance through the experiment phase. AAF also provides deployment support with automated CI/CD processes. The resulting system is intended for not only enhancing efficiency and scalability but also achieve quality decision making and inference with optimized cost and timeline compared to manual methods.