OPTAgent: From Standalone Solvers to Deployment-Aware Solver Ecosystems

Much of numerical optimization research develops and evaluates individual algorithms or solver components within fixed problem settings. As a result, continued advances in numerical optimization have steadily expanded the toolbox of potential solver methods, configurations, and complementary procedures. This has made it increasingly difficult to identify which method to use on a particular problem instance. Moreover, available computational resources add an important dimension to method selection and configuration, in particular when many problems must be solved in a limited amount of time using shared resources. To address this challenge, we introduce OPTAgent, an agentic policy-synthesis framework that leverages LLMs to construct method selection and configuration policies from a toolbox of optimization solvers. During offline discovery, an LLM proposes and searches a space of compact decision-tree policies that allocates a computational budget to method(s) and their configurations to solve collections of problems.