Agent for Robotaxi Safety Policy Analysis & Dynamic Procedure Generation
Develop an advanced agent using Anthropic's Claude Agents SDK to analyze real-time robotaxi incident data, cross-reference it with complex safety regulations, and dynamically generate or adapt operational procedures. Inspired by the Waymo incident and upcoming UK regulations, this challenge focuses on building a highly reliable, safety-critical agent that can interpret regulatory documents, learn from incidents, and output actionable safety protocols. The agent will leverage Claude Opus's extended thinking and 'computer use' capabilities to process vast amounts of unstructured text and adapt to evolving regulatory landscapes.
What you are building
The core problem, expected build, and operating context for this challenge.
Build a Claude Agent SDK-based system with Claude Opus 4.1 for real-time robotaxi incident analysis, safety policy interpretation, and dynamic procedure generation.
How work is evaluated
The evaluation will assess the Claude agent's ability to accurately analyze robotaxi incidents, correctly interpret safety policies, and generate effective, compliant, and safe operational procedures. It will emphasize the agent's reasoning, 'computer use' for document interaction, and the quality of its output.
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.
PolicyReferenceCorrectness
Verifies that generated procedures correctly reference the relevant safety policies or regulations.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
ProcedureSafetyStandard
Checks if the generated procedure includes specific, quantifiable safety improvements.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
ReasoningDepthScore
A subjective score (1-5) on the depth of the agent's root cause analysis and understanding of policy implications. • target: 4 • range: 1-5
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
ProcedureSpecificityScore
A score (1-5) on how specific, actionable, and unambiguous the generated procedures are. • target: 4 • range: 1-5
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Master the Claude Agents SDK for defining agent capabilities, tools, and orchestrating complex multi-step reasoning processes.
Utilize Claude Opus 4.1 for advanced text comprehension, policy interpretation, and generating nuanced operational procedures.
Implement 'computer use' tools within the Claude Agent for tasks such as simulating incident scenarios, extracting specific clauses from regulatory PDFs, and querying external knowledge bases.
Develop custom tools for the Claude Agent to interface with a simulated real-time incident stream and a database of safety policies.
Design a robust AI evaluation harness (e.g., using LangChain Eval or custom Python scripts) to assess the correctness, completeness, and safety of generated procedures.
Orchestrate the agent's decision-making to dynamically adapt safety protocols based on new incident data or updated regulations, ensuring compliance and continuous improvement.
Reference links and supporting material
Sample data for 2 tasks
Challenge setup
The evaluation will assess the Claude agent's ability to accurately analyze robotaxi incidents, correctly interpret safety policies, and generate effective, compliant, and safe operational procedures. It will emphasiz...
Challenge input
{ "incident_report": "string", "relevant_policies": "array" }
Agent execution
No agent tool is listed for this challenge.
Evaluated output
{ "analysis_summary": "string", "root_causes_identified": "array", "generated_procedure": "string", "policy_references": "array" }
- Verifies that generated procedures correctly reference the relevant safety policies or regulations.
- Checks if the generated procedure includes specific, quantifiable safety improvements.
- A subjective score (1-5) on the depth of the agent's root cause analysis and understanding of policy implic...
- ReasoningDepthScore target: 4
- 2 public reference cases
- Python execution harness
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
No tool records are attached.
Evaluation contract
- PolicyReferenceCorrectness · Weight 1
- ProcedureSafetyStandard · Weight 1
- ReasoningDepthScore · Weight 1
- ProcedureSpecificityScore · Weight 1
Metrics
- ReasoningDepthScore · Target 4
- ProcedureSpecificityScore · Target 4
Reference cases
- 2 public reference cases
Workflow tasks
- Task 1
Input: { "incident_report": "string", "relevant_policies": "array" }
Output: { "analysis_summary": "string", "root_causes_identified": "array", "generated_procedure": "string", "policy_references": "array" }
Execution harness: Python
- Task 2
Input: { "new_regulation_text": "string", "existing_procedure_text": "string" }
Output: { "compliance_status": "string", "identified_gaps": "array", "recommended_updates": "string" }
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.