How ibm/assetopsbench works
AssetOpsBench is a framework for industrial asset-operations agents, MCP domain tools, and offline evaluation. The primary workflow is an engineer’s question handled by an agent, which uses domain servers to retrieve asset, work-order, failure-mode, or time-series evidence and return an operational answer. Distinct branches provide TSFM forecasting and catalog workflows, vibration analysis, and evaluation/reporting. The graph reflects documented capabilities and indexed imports; unsampled implementation details are not assumed, so uncertain relationships are omitted.
Main entry points: agent/runner.py Agent runner · main.py TSFM tool server · composition.py Recipe composition · evaluation/runner.py Evaluation runner
Components 32
Agent orchestration
- Agent runner
src/agent/runner.py - Plan-execute
src/agent/plan_execute/executor.py - Deep Agent
src/agent/deep_agent/runner.py - Claude Agent
src/agent/claude_agent/runner.py - OpenAI Agent
src/agent/openai_agent/runner.py - OpenCode Agent
src/agent/opencode_agent/runner.py - Stirrup Agent
src/agent/stirrup_agent/runner.py
Model backends
- LLM backend
src/llm/base.py - Backend selection
src/llm/__init__.py - Proxy routing
src/llm/routers.py
Industrial MCP tools
- IoT assets and sensors
src/servers/iot/main.py - Failure modes
src/servers/fmsr/main.py - Work orders
src/servers/wo/main.py - Work-order lifecycle
src/servers/wo/workorders.py - Vibration diagnostics
src/servers/vibration/main.py - Signal analysis
src/servers/vibration/dsp/fault_detection.py - Shared utilities
src/servers/utilities/main.py
Time-series analysis
- TSFM tool server
src/servers/tsfm/main.py - Recipe composition
src/servers/tsfm/engine/composition.py - Feature reasoning
src/servers/tsfm/reasoning/feature_selection.py - Model catalog
src/servers/tsfm/stores/model_store.py - Feature catalog
src/servers/tsfm/stores/feature_store.py - Model resolver
src/servers/tsfm/substrate/resolver.py
Evaluation and telemetry
- Evaluation runner
src/evaluation/runner.py - Scenario scorers
src/evaluation/scorers/static_json.py - LLM judge
src/evaluation/scorers/llm_judge.py - Metrics
src/evaluation/metrics.py - Reports
src/evaluation/report.py - Tracing and persistence
src/observability/tracing.py
Other components
- Maintenance engineer
- CouchDB
- MCP tool environment
Connections 38
- Maintenance engineer → Agent runner: asks task
README.md - Agent runner → Plan-execute: selects framework
README.md - Agent runner → Deep Agent: selects framework
README.md - Agent runner → Claude Agent: selects framework
README.md - Agent runner → OpenAI Agent: selects framework
README.md - Agent runner → OpenCode Agent: selects framework
src/agent/opencode_agent/runner.py - Agent runner → Stirrup Agent: selects framework
src/agent/stirrup_agent/runner.py - Agent runner → Backend selection: creates backend
src/agent/runner.py - Backend selection → LLM backend: selects implementation
src/llm/__init__.py - Backend selection → Proxy routing: checks prefix
src/llm/__init__.py - Agent runner → MCP tool environment: uses tool servers
README.md - MCP tool environment → IoT assets and sensors: exposes tools
README.md - MCP tool environment → Failure modes: exposes tools
README.md - MCP tool environment → Work orders: exposes tools
README.md - MCP tool environment → Vibration diagnostics: exposes tools
README.md - MCP tool environment → TSFM tool server: exposes tools
README.md - MCP tool environment → Shared utilities: exposes tools
README.md - Work orders → Work-order lifecycle: delegates lifecycle
src/servers/wo/main.py - Work orders → CouchDB: reads and writes
src/servers/wo/main.py - IoT assets and sensors → CouchDB: queries asset data
src/servers/iot/main.py - Vibration diagnostics → Signal analysis: runs diagnostics
src/servers/vibration/main.py - Vibration diagnostics → CouchDB: loads sensor data
src/servers/vibration/main.py - TSFM tool server → Recipe composition: dispatches recipes
src/servers/tsfm/main.py - TSFM tool server → Feature reasoning: uses feature selection
src/servers/tsfm/main.py - TSFM tool server → Model catalog: looks up models
src/servers/tsfm/main.py - TSFM tool server → Feature catalog: looks up features
src/servers/tsfm/main.py - TSFM tool server → Model resolver: resolves models
src/servers/tsfm/main.py - Recipe composition → Model catalog: loads model cards
src/servers/tsfm/engine/composition.py - Recipe composition → Feature catalog: loads feature cards
src/servers/tsfm/engine/composition.py - Recipe composition → Model resolver: resolves model
src/servers/tsfm/engine/composition.py - TSFM tool server → CouchDB: uses catalog data
src/servers/tsfm/main.py - Evaluation runner → Scenario scorers: scores scenarios
- Evaluation runner → Metrics: calculates metrics
- Evaluation runner → Reports: produces report
- Scenario scorers → LLM judge: offers judge scorer
- LLM judge → LLM backend: requests judgment
src/evaluation/scorers/llm_judge.py - Reports → Metrics: summarizes metrics
src/evaluation/report.py - Tracing and persistence → CouchDB: exports traces