AWS re:Invent 2025 - Implementing Human-in-the-Loop Controls for Multi-Agent AI Systems (CNS428)

Implementing Human-in-the-Loop Controls for Multi-Agent AI Systems

Evolution of AI Assistants to Autonomous Agents

• Progression from simple generative AI assistants (chatbots) to more advanced "agentic AI systems"
• Agentic AI systems can work behind the scenes, not just through a chatbot interface
• Increasing autonomy of systems, but need for human oversight and ownership

AWS Progression in Model Inferencing and Contextual Responses

• Early options: SageMaker, ECS/EC2, EKS for running models
• Evolution to include persistent stores for chat history and vector stores for additional context
• Moving towards real-time data integration using agents and tools

Securing and Observing Agent-Based Systems

• Need for inbound/outbound authentication, observability, and guardrails when building agent-based systems
• Integration with MCP (Multi-Agent Coordination Protocol) servers and A2A (Agent-to-Agent) communication

Determining When Human-in-the-Loop (HITL) is Necessary

• High-stake decision points (e.g., prescribing medicine)
• Irreversible actions (e.g., financial transactions)
• Regulatory compliance requirements
• Building trust in new agentic AI systems

Implementing HITL Controls

• Using MCP elicitations for interactive user input
• Leveraging Step Functions "wait for callback" functionality
• Integrating HITL controls in tools like Langraph

Reducing HITL as Autonomy Increases

• Continuous agent evaluation (e.g., Agent Core evaluations) to measure autonomy
• Transitioning towards a "centaur system" - symbiotic human-AI collaboration

Key Takeaways and Resources

• HITL is crucial for high-stakes decisions, irreversible actions, and building trust in new agentic AI systems
• Implement HITL controls using MCP elicitations, Step Functions, and integration with tools like Langraph
• Gradually reduce HITL as agent autonomy and evaluation matures, moving towards a "centaur system"
• Refer to resources like the Harvard paper on "human-algorithm centaurs" and the sample application using A2A