Skip to content

AWS re:Invent 2025 - AI agents in action: Architecting the future of applications (INV202)

This presentation by Shaown Nandi, Director of Technology for AWS, explores how agentic AI is transforming cloud-native application architecture and unlocking significant value for businesses (1:24).

1. The Agentic Era and Its Transformative Impact (3:00)

Problem

Traditional AI implementations have delivered only incremental gains through simple chatbot integrations, failing to unlock the exponential value potential of artificial intelligence.

Solution

The Agentic Advantage: Agentic AI is positioned to deliver exponential value beyond incremental improvements, including significant cost savings, increased productivity, and enhanced efficiency (3:44).

  • Amazon's Internal Success: Over $2 billion in annual cost savings and 4.5x increase in development team productivity (4:26 - 4:58)
  • Customer Success Stories:
  • ASAP: Automated 90% of end-user interactions, achieving 91% first-call resolution while reducing costs by 77% (5:15)
  • NovacMP: 60% reduction in technical debt, transforming 3-week modernization tasks to under an hour (5:34)
  • No Harm: 8x boost in patient capacity with over $30 million in annual savings (5:54)

Outcome

Organizations are experiencing fundamental shifts in business value calculus, with significantly decreased barriers to entry for AI implementation (7:20).

2. Architectural Evolution: From Static to Dynamic Logic (8:41)

Problem

Traditional applications rely on static, code-defined logic that cannot adapt dynamically to changing contexts or requirements.

Solution

Dynamic LLM-Augmented Systems: Agentic systems move logic into LLM-augmented agents that can interpret context, make decisions, and adapt dynamically (9:12 - 9:33).

  • Flexible Patterns: Distributed system architecture requiring less engineering complexity to integrate LLMs into business processes (6:45)
  • Overcoming Early GenAI Limitations: Early architectures were rigid and difficult to integrate into production systems (10:22)
  • Enhanced Capabilities: Models can maintain goal-driven loops, retain context, integrate tools natively, and operate within scalable environments (11:28)

Outcome

Shift from deterministic to dynamically interpreted systems, enabling agents to perform tasks at scales and speeds impossible for humans (8:09 - 9:53).

3. AWS Tools for Agent Development (13:09)

Problem

Building production-ready agents requires significant engineering effort, often taking months to implement goal-driven loops, planning, and operational concerns.

Solution

Comprehensive Development Platform:

  • Strands: Open-source framework handling goal-driven loops, planning, acting, and reflecting cycles while abstracting security and observability (13:09)
  • Amazon Bedrock Agent Core: Managed, enterprise-grade runtime platform providing session isolation, long-running workload support, tool integration, memory services, and policy enforcement (14:12)

Outcome

Engineering effort reduced from months to hours, enabling rapid agent development and deployment (14:50).

4. Customer Implementation: Steady Healthcare (16:00)

Problem

Healthcare eligibility checks frequently fail due to misspelled names or missing information, leading to rejections, phone calls, and delayed care (17:30).

Solution

Rapid Agentic Implementation: Built first agentic feature in two weeks using Agent Core and Strands, focusing on a narrow, low-risk use case (18:42).

  • Security Approach: Principle of least privilege with agents having only necessary API access, layered with Bedrock guardrails (28:04)

Outcome

  • Adoption: Nearly one-third of customers using the feature weekly (32:35)
  • Resolution Rate: Nearly one-third resolution rate for eligibility checks, avoiding phone calls and denied care instances
  • Expansion Plans: Extending to payer selection and claims processing (33:48)

5. Customer Implementation: QAD Manufacturing (19:23)

Problem

Manufacturing systems traditionally function as systems of record rather than systems of action, limiting employee and machine capability to drive outcomes.

Solution

Three-Bucket Agent Strategy (20:27):

  • Persona-Based Agents: Automate mundane tasks for various roles (21:41)
  • Optimization Agents: Handle complex problems like inventory carrying costs and procurement (22:23)
  • Implementation Agents: Drastically reduce system deployment times (22:58)

Security by Design: Ensuring autonomy doesn't decrease control through AWS controls like IAM, audit trails, and KMS (29:48), with "trust but verify" approach keeping humans in the loop for critical decisions (30:35).

Outcome

Measurable Business Impact: - Red Zone Solution: 26% productivity increase, 81% more engagement, 35% reduction in attrition (31:10) - Inventory Agent: Up to 30% reduction in carrying costs (32:01) - Procurement Agent: 50% reduction in buyer time (32:01)

6. The Future of Agentic AI (35:56)

Problem

Current business processes are constrained by sequential handoffs, limited planning horizons, and reactive quality checks.

Solution

Four Key Capabilities:

  • Accelerated Execution: Agents collapse handoffs, plan multiple steps ahead, and execute work in parallel (35:56)
  • Adaptive Orchestration: Work from intent, explore alternative paths, test hypotheses, and dynamically reason about decisions (36:52)
  • Embedded Quality & Resilience: Quality checks happen during work performance, with agents validating intermediate results, self-correcting, and retrying failed steps (37:27)
  • Scalable Reach & Persistent Memory: Pull context across data silos, reason over various data types, and remember successful approaches (38:04)

Outcome

Industry Transformation: Evolution from LLM-assisted workflows to fully agentic processes delivering extraordinary results, lower costs, and faster cycle times (39:38). Amazon has deployed over 20,000 agents since July (39:56).

7. AI-Native Development Revolution (40:12)

Problem

Traditional software development involves sequential processes with delayed feedback and siloed team interactions.

Solution

Agentic Development Orchestration: Agents transform software development by orchestrating the entire lifecycle, bringing different teams together in real-time, and surfacing trade-offs early (40:12).

Outcome

Accelerated Development Examples: - Repro: Developed a healthcare platform in 20 hours (41:08) - DAN: Built new capabilities in 48 hours (41:08) - Pattern E-commerce: Reimagined business processes with agents at the core (41:55 - 42:00)

Presenter: Shaown Nandi - Director of Technology at AWS