AssetOpsBench

AssetOpsBench, published under IBM's GitHub organization, is a framework for developing, orchestrating and evaluating domain-specific AI agents for industrial asset operations and maintenance. It provides reproducible scenarios, agent tooling and evaluation pipelines for multi-step workflows in simulated industrial environments. The README's summary lists 9 asset classes, 141+ scenarios and 5 domain agents, and says the work was accepted to KDD 2026 (Datasets & Benchmarks track). Example tasks include listing the sensors of a chiller at a site (IoT) and identifying failure modes.

Domain-specific MCP servers expose the tools: IoT (sites, assets, sensors, history), failure-mode and sensor mapping (FMSR), time-series foundation models (forecasting and anomaly detection), work orders and vibration analysis (FFT and envelope spectra). Agent frameworks include Plan-Execute, Deep Agent, and ReAct orchestrators for Claude and OpenAI models. Setup is `git clone` and `pip install -e .`, with a Colab notebook and a Hugging Face playground for trying it without installing, and INSTRUCTIONS.md for MCP servers and the plan-execute runner. The README notes that active development is on `main`, while publication code stays on separate branches. It is Apache-2.0 licensed, with a paper and dataset linked.