Ethical Agent for Adaptive User Safety & MCP Policy
This challenge focuses on building a proactive ethical AI agent system. You will use Langroid to construct a robust, stateful agent capable of monitoring user interactions in real-time, coupled with Smolagents for reactive and lightweight responses. Claude Sonnet 4 will be central to the agent's ability to understand nuanced user sentiment and potential mental health risks. The system must implement age-gated policies and usage limits by dynamically integrating MCP for policy enforcement and leveraging adaptive thinking budgets to determine the appropriate level of intervention or support, including deploying hybrid instant/deep reasoning to balance immediate safety actions with comprehensive ethical analysis. Guidance will be used to ensure structured, safe conversational outputs.
What you are building
Build an ethical AI agent using Langroid, Claude Sonnet 4, and Smolagents to enforce user safety policies, adapt to mental health cues, and integrate dynamic MCP policy rules with adaptive thinking budgets.
How work is evaluated
The evaluation will assess the agent's ability to accurately detect risks, adhere to defined policies (age-gating, time limits), and generate safe, appropriate responses in simulated chat scenarios.
Datasets
Public datasets, plus the private datasets you have access to.
Learning goals
Deploy hybrid instant/deep reasoning systems for immediate safety actions and thorough ethical analysis.
Master Langroid for developing robust, stateful ethical AI agents.
Implement Smolagents for real-time, reactive monitoring of user interactions.
Apply adaptive thinking budgets with Claude Sonnet 4 for dynamic risk assessment and intervention strategies.
Design MCP-enabled policy enforcement tools for age-gating, content moderation, and usage limits.
Integrate Guidance for ensuring structured, safe, and policy-compliant conversational outputs.
Resources
Sample data for 3 tasks
Challenge setup
The evaluation will assess the agent's ability to accurately detect risks, adhere to defined policies (age-gating, time limits), and generate safe, appropriate responses in simulated chat scenarios.
Challenge input
{'user_profile': {'age': 16, 'chat_duration_minutes': 110}, 'user_message': 'string'}
Agent execution
No agent tool is listed for this challenge.
Evaluated output
{'policy_violation_detected': bool, 'agent_action': 'string', 'policy_type': 'age_gate'|'time_limit'|'none'}
- Verify that agent state (e.g., chat duration) is correctly maintained across turns.
- Ensure the agent successfully retrieves policy rules via MCP.
- Confirm that structured responses correctly apply predefined Guidance templates.
- Policy Adherence Rate target: 90%
- Python execution harness
Shows the first task, up to three tools, and up to three evidence items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- Verify that agent state (e.g., chat duration) is correctly maintained across turns.
- Ensure the agent successfully retrieves policy rules via MCP.
- Confirm that structured responses correctly apply predefined Guidance templates.
Metrics
- Policy Adherence Rate · Target 0.9
- Mental Health Detection Recall · Target 0.8
- Response Safety Score · Target 4.5
Workflow tasks
- Task 1
Input: {'user_profile': {'age': 16, 'chat_duration_minutes': 110}, 'user_message': 'string'}
Output: {'policy_violation_detected': bool, 'agent_action': 'string', 'policy_type': 'age_gate'|'time_limit'|'none'}
Execution harness: Python
- Task 2
Input: {'chat_history': ['message1', 'message2', ...], 'user_id': 'string'}
Output: {'risk_level': 'low'|'medium'|'high', 'intervention_recommendation': 'string', 'confidence': float}
Execution harness: Python
- Task 3
Input: {'user_message': 'string', 'context': 'string', 'risk_detected': bool}
Output: {'safe_response': 'string', 'guidance_template_used': 'string', 'ethical_compliance_score': float}
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.