Specialized LM Agents with Simulation-Verified Search for Service Workforce Planning

Operational planning in industrial field service requires jointly optimizing heterogeneous decisions -- shift scheduling, field assignments, and workforce skills management, while balancing competing KPIs. We investigate whether assigning distinct action types to specialized LM agents produces better plans than a single monolithic LM reasoning over all actions jointly. Both architectures generate structured proposals that are evaluated via Monte Carlo Tree Search against a high-fidelity discrete-event simulation initialized from real operational data. Experiments on real-world service operations across multiple regions show that (i) specialized agents produce higher-quality proposals than a monolithic LM under equal search budgets, (ii) simulation-based search substantially improves over greedy LM-confidence-based selection, and (iii) LM-generated proposals outperform random candidates, confirming that the LM encodes useful domain structure.