AssetOpsBench is an open-source framework developed by IBM for building, orchestrating, and evaluating domain-specific AI agents in Industry 4.0 asset operations and maintenance. Written in Python, the repository provides a unified platform that combines large language models with time series foundation models to address real-world industrial challenges across maintenance engineering, reliability, and facility planning.
The framework encompasses 460 plus scenarios distributed across 9 asset classes, with 5 specialist agents covering IoT data access, failure mode and sensor relationships (FMSR), time series forecasting (TSFM), work order management, and vibration analysis. These agents operate through domain-specific MCP (Model Context Protocol) servers that expose critical tools for querying sensor data, retrieving failure modes, performing forecasting and anomaly detection, and managing work orders. The repository includes multiple agent frameworks for different use cases: a plan-execute sequential workflow compatible with any LLM, a deep agent supporting planning and sub-agents for long-horizon tasks, and ReAct-based orchestrators for Claude and OpenAI models with agent-as-tool delegation.
Multi-agent orchestration is supported through blueprints including MetaAgent and AgentHive, enabling complex workflows that coordinate multiple specialist agents over live sensor data and Industry 4.0 records such as FMEA documents and maintenance alerts. The framework provides reproducible evaluation pipelines with leaderboards scoring agent trajectories across 6-dimensional criteria measuring reasoning, execution, and data handling capabilities.
The project has achieved significant academic recognition, with acceptance at KDD 2026 for the Datasets and Benchmarks track alongside a hands-on tutorial on building reliable industrial agents with MCP. Additional publications span 12 plus contributions across 7 top-tier venues in 2025 and 2026, including work at ACL, ICLR, AAAI, IAAI, NeurIPS, and EMNLP. The framework powers public AI competitions including the IJCAI 2026 Industrial Automation Challenge focused on physics-grounded LLM reasoning.
The repository provides multiple entry points for users: a Colab notebook for immediate experimentation, a Hugging Face playground for interactive exploration, and comprehensive documentation in INSTRUCTIONS.md. Active development occurs on the main branch while publication-specific implementations are maintained on separate branches such as IndustryAssetEQA for ACL 2026 work and main-0.x for earlier experimental work. The associated dataset is available on Hugging Face, and the framework has generated over 500 competition submissions from researchers and practitioners worldwide.
