AWS re:Invent 2025 - AgenticAI: Industry-Driven Architecture Excellence (NTA203)

Summary of AWS re:Invent 2025 - AgenticAI: Industry-Driven Architecture Excellence (NTA203)

Introduction to Agentic AI

• Agentic AI is an evolution in the AI space, complementary to but distinct from Generative AI
• Agentic AI is focused on completing specific tasks and solving complex problems, often involving subjective outcomes and large amounts of data
• Agentic AI is being rapidly adopted across industries to solve business challenges and create transformative customer experiences

Agentic AI in Healthcare

• Generative AI has been used for chatbots, care management, clinical trial matching, and medical research
• Agentic AI is enabling intelligent patient experiences, autonomous clinical workflows, and supply chain optimization
• Agentic AI uses an orchestration layer to coordinate agents across multiple domains, enabling a shift from reactive to proactive healthcare operations

Architectural Guidance for Agentic AI

• Key functions of an agent: Reasoning using LLMs, goal-setting, access to tools and enterprise data, observability and guardrails
• Reactive vs. Proactive Agents: Using multiple specialized agents in a composite architecture to address complex use cases
• Agent-to-Agent Communication: Using an open standard protocol for interoperability and transport-agnostic communication
• Architectural Patterns:
  • Agents as Tools: Orchestrator agent using other specialized agents as callable functions
  • Dynamic Customer Pricing: Agents deciding the communication path based on customer segmentation
  • Peer-to-Peer Agents: Agents sharing a common knowledge base and memory to make collaborative decisions
  • Workflow Pattern: Deterministic, sequential flow of agents performing specific tasks

Path to Production

• Start with a business problem and build a value model to justify the investment
• Key cost elements: Model API access, optimization, infrastructure, data strategy, talent/development, operations and support, ethical AI
• Potential revenue streams: Direct (per-access, subscription) and indirect (increased sales, customer experience)
• Production challenges: Performance, scalability, security, context, governance, agent silos
• Production architecture:
  • Application layer: Web/mobile apps, chat interfaces, authentication
  • Agent layer: Agents and MCP services, accessing LLMs via AWS Bedrock
  • Core layer: Agent Core for deployment, observability, and security
  • Semantic layer: Data transformation and connectivity using AWS Glue and Neptune
  • Data layer: Leveraging existing data sources, no need for a full data lake

Innovations in the AI Development Lifecycle (AIDL)

• AIDL is a methodology to transform the traditional software development lifecycle using AI
• Key phases:
  • Inception: Using AI agents to break down the high-level intent into specific units of work
  • Construction: Prompting AI to write code and tests, with humans guiding and reviewing
  • Operations: Leveraging AI for infrastructure-as-code and deployment
• Benefits: Demonstrated 2-month backlog delivered in 2 days using AIDL, compared to traditional SDLC

Conclusion

• Agentic AI presents significant opportunities across industries to solve complex problems and create transformative experiences
• Successful implementation requires a strategic approach, considering value models, architectural patterns, and a shift in the development lifecycle
• AWS is working to enable organizations to reimagine their business with secure, reliable, and scalable Agentic AI solutions