MeshAgent: Enabling Reliable Network Management with Large Language Models

Large language models (LLMs) are promising for domain-specific agents, but adapting them to network management remains difficult because existing specialization methods depend on large amounts of high-quality task data, while real network queries are diverse and unpredictable. To address this, we propose MeshAgent, a workflow that extracts reusable domain invariants from sample queries and encodes them as constraints to guide LLM generation and validation. Across three network management applications and a user study with industry professionals, MeshAgent consistently improves accuracy and reliability, including the ability to abstain when confidence is low. MeshAgent achieves over 95% accuracy, reaches 100% when combined with fine-tuned agents, and improves accuracy by up to 26% over baselines. These results show that constraint-based adaptation is a practical and scalable alternative to data-heavy LLM specialization for network management.