From Portfolio Solvers to Agentic Solvers

This paper explores the shift from traditional portfolio solvers to agentic solvers, enabled by Large Language Models (LLMs) to support autonomous and collaborative behavior across solving systems. Beyond per-instance algorithm selection, agentic solvers can dynamically adapt strategies, refine solutions, and reformulate models. LLMs act as orchestrators, enhancing flexibility, scalability, and adaptivity. This work was submitted on 21 June 2025 and accepted on 25 June 2025 for the LLM-Solve Workshop at CP/SAT/SoCS 2025.