A Vision for a Logic-Based Workflow Formulation Framework and an Agentic Execution Pipeline for Pattern Engineering
Organizations increasingly depend on insights drawn from data to inform consequential decisions. Many concern structural behaviors, which we call patterns. Producing them may require working with multiple pattern types, performing different operations over them, and composing those operations into workflows, an activity we call pattern engineering. No principled framework exists for organizing such workflows: analysts devise each by hand, and existing systems automate particular analytical objectives rather than representing pattern operations in a general, composable form. We introduce a typed framework that represents each pattern through a fixed set of modular components, and expresses pattern operations as queries which specify input and output components. Chaining queries composes workflows, while execution of any individual query remains unconstrained. We instantiate the framework with an LLM-based agentic executor and evaluate example-mining, example-alignment, and pattern-induction queries over two temporal patterns, using World Intellectual Property Organization (WIPO) reports and patent data. All three queries are carried out across four LLM settings, which perform differently by target component, while the traces allow us to localize errors by execution stage.