AWS re:Invent 2025 - Using AI to improve humanitarian workload resilience (AIM336)

Leveraging AI to Enhance Workload Resilience

Importance of Resilient Humanitarian Workloads

• Humanitarian workloads, designed to help people in need during crises, require high reliability and resilience
• Failures like timeouts or 404 errors are unacceptable when people are calling for emergency assistance
• Building resilient systems is challenging, requiring attention to high availability, disaster recovery, and continuous improvement

Key Resilience Concepts

• High Availability: Ability to withstand common failures like server outages or network interruptions
• Disaster Recovery: Ability to fail over to a secondary site and recover from major disruptions
• Continuous Improvement: Iteratively enhancing resilience over time based on learnings from failures

Resilience Failure Categories

1. Single Points of Failure: Single components or systems that, if they fail, bring down the entire application
2. Excessive Load: Inability to scale and handle sudden increases in traffic
3. Excessive Latency: Timeouts and failures due to unpredictable network latency
4. Misconfiguration and Bugs: Issues caused by code defects or configuration errors
5. Shared Fate: Common dependencies that, when they fail, impact multiple applications

Leveraging Agentic AI

• Building an AI agent with agency to autonomously investigate and reason about workload resilience
• Utilizing Amazon Bedrock foundation model and existing AWS tools like Cloud Control API and AWS documentation

Agent Architecture and Functionality

1. Foundation Model: Using Amazon Bedrock for reasoning and inference
2. Cloud Control API: Interacting with AWS resources to assess the current state
3. AWS Documentation MCP: Retrieving relevant documentation to provide recommendations

Detailed Prompt Engineering

• Separating the agent's prompt into a separate file for easier maintenance and iteration
• Defining a comprehensive prompt covering:
  • Agent persona and expertise
  • Grading philosophy and scale
  • Resilience definitions and requirements
  • Specific instructions for tool usage and output formatting

Demonstration and Results

• Agent analyzes a "food agent" workload with 24-hour RTO and 12-hour RPO
• Provides detailed resilience assessments, grading the workload across key areas
• Adjusts recommendations when RTO/RPO requirements change
• Identifies observability improvements and generates a recovery runbook

Key Takeaways

• Agentic AI can significantly simplify the process of building resilient systems by automating analysis and providing actionable guidance
• Comprehensive prompt engineering is crucial to ensure the agent provides consistent, reliable, and contextually appropriate recommendations
• Integrating the agent with existing AWS tools and documentation enables it to leverage a wealth of resources to support its analysis
• Deploying the agent on a platform like Bedrock Agent Core can make it accessible and usable across an organization

Next Steps and Resources

• Explore the Strands Agent documentation for more information on building AI-powered agents
• Check out the nonprofit sample code on GitHub to get hands-on with the resilience agent
• Reach out to your AWS account team for assistance in implementing similar solutions
• Utilize AWS Skill Builder resources for further learning on resilience, AI, and AWS services