Which two validation approaches are MOST critical for ensuring agent reliability in production deployments?
(Choose two.)

• A. Performance testing during development phases
• B. Random sampling of agent interactions for manual review
• C. User satisfaction surveys as the primary quality metric
• D. Automated consistency checking across multiple agent runs
• E. Structured output validation with Pydantic schemas

A Lead AI Architect at a global financial institution is designing a multi-agent fraud detection system using an agentic AI framework. The system must operate in real time, with distinct agents working collaboratively to monitor and analyze transactional patterns across accounts, retain and share contextual information over time, and escalate suspicious behaviors to a human fraud analyst when needed.
Which architectural approach enables intelligent specialization, shared memory, and inter-agent coordination in a dynamic and evolving threat environment?

• A. Design an agentic workflow where each agent acts independently on isolated data slices with no inter- agent communication to reduce latency and model complexity.
• B. Design a multi-agent system where individual agents collaborate synchronously using shared memory and structured messaging.
• C. Design monolithic LLM-based agents that handle all fraud detection tasks within a single loop, without modular roles or multi-agent coordination.
• D. Design a modular multi-agent system where individual agents collaborate asynchronously using shared memory and structured messaging.
• E. Design a centralized rule-based service that checks all transactions against static fraud indicators and sends alerts when thresholds are exceeded.

In a production agentic system handling thousands of concurrent conversations, which state management strategy provides optimal performance while ensuring context preservation?

• A. Global shared state with locks for concurrent access
• B. Session-isolated state with serialization and lazy loading
• C. Stateless design with context reconstruction from message history

You are deploying an AI-driven applicant-screening agent that analyzes candidate resumes and social-media data to recommend top applicants. Due to anti-discrimination laws and corporate policy, the system must mitigate bias against protected groups, maintain an audit trail of decisions, and comply with GDPR (including data minimization and explicit consent).
Which of the following strategies is most effective for ensuring your screening agent both mitigates bias in its recommendations and complies with data-privacy regulations?

• A. Exclude gender and ethnicity fields during training, use a generic privacy policy for consent, and do not maintain audit logs or apply targeted debiasing.
• B. Perform a post-deployment GDPR and bias audit and process raw personal data as received.
• C. Encrypt all candidate data at rest and in transit, remove protected attributes from analysis, and conduct manual bias checks on recommendations.
• D. Pseudonymize protected attributes, implement fairness-aware debiasing, maintain an audit trail, and enforce GDPR data-minimization and consent.