Autonomous Factory Troubleshooting Agent with Claude Agents SDK
When critical automation lines stall, control engineers must navigate multiple HMI screens, review real-time sensor metrics, and consult deep reasoning models to resolve root causes. In this challenge, developers will utilize Anthropic's Claude Agents SDK featuring extended thinking and computer use capabilities to build an autonomous factory troubleshooting assistant. The Claude Agent uses extended thinking to systematically formulate diagnostic hypotheses for sudden pH or vibration anomalies, inspect virtual HMI screen interfaces using computer use tools, and issue structured command sequences to stabilize plant operations.
What you are building
The core problem, expected build, and operating context for this challenge.
Develop an autonomous factory troubleshooting assistant using Claude Agents SDK with extended thinking and tool execution capabilities.
How work is evaluated
Evaluates diagnostic accuracy and logical completeness of extended thinking steps during plant fault simulation.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
Cavitation Diagnosis Verification
Correctly diagnoses filter blockage and issues BACKWASH command
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Thinking Trace Depth Score
Evaluation of multi-step logical deduction steps in extended thinking output • target: 0.95 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Implement autonomous reasoning loops using Claude Agents SDK
Enable extended thinking capabilities for deep multi-step root-cause analysis
Create custom tools for simulated HMI interaction and PLC command execution
Generate verifiable step-by-step resolution logs for industrial audit trails
Reference links and supporting material
Simulated plant alarm sequences and HMI UI layout snapshots representing hydraulic press and fluid loop failures.
Challenge setup
Evaluates diagnostic accuracy and logical completeness of extended thinking steps during plant fault simulation.
Challenge input
JSON containing plant_alarm_event and sensor_telemetry_snapshot
Claude Agents SDK
Assigned primary orchestration SDK
Evaluated output
JSON containing reasoning_summary, target_plc_command, safe_execution_flag (boolean)
- Correctly diagnoses filter blockage and issues BACKWASH command
- Evaluation of multi-step logical deduction steps in extended thinking output • target: 0.95 • range: 0-1
- Benchmark: Complex Fault Diagnosis Benchmark
- Thinking Trace Depth Score target: 0.95
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Claude Code · Required
- Anthropic · Optional
- Claude Code · Required
- Claude 4.1 Opus · Optional
Evaluation contract
- Cavitation Diagnosis Verification · Weight 1
- Thinking Trace Depth Score · Weight 1
Metrics
- Thinking Trace Depth Score · Target 0.95
Benchmarks
- Complex Fault Diagnosis Benchmark
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing plant_alarm_event and sensor_telemetry_snapshot
Output: JSON containing reasoning_summary, target_plc_command, safe_execution_flag (boolean)
Execution harness: Python
Evaluation tools
- Claude Agents SDK · Reasoning agent SDK
Recipe freeze marker
No recipe freeze time is recorded.