Pipeline-agent: A comprehensive decision support system for natural gas pipelines via verifiable computational reasoning

Natural gas transmission networks require rapid operational decisions that integrate geographic information systems data, network topology, and time indexed measurements from operational systems. While large language models can reduce the interaction barrier for engineers, text only generation is unreliable for safety critical computations and rarely provides auditable evidence. This paper presents Pipeline-Agent, a decision support system for natural gas pipelines built on verifiable computational reasoning. Pipeline-Agent aligns heterogeneous sources into a unified spatiotemporal schema, creates an explicit plan, executes only deterministic tools within a sandboxed workspace, and returns a verifiable result package that includes intermediate artifacts and execution logs. We further introduce Pipeline-Bench, a benchmark with eighty task templates spanning eight categories. Across evaluation runs, Pipeline-Agent achieves a 2.24 × higher mean score than the best baseline and a 2.56 × higher mean score than OpenAI GPT-5.2 direct generation. The results indicate that tool grounded agents with explicit planning and verification can provide practical and transparent decision support for pipeline operations.