The Agentic AI Era Seen at AWS re:Invent 2025

AWS re:Invent 2025 was a stage that showed Agentic AI has entered the production phase. Kakao Bank engineers looked at the new announcements from the perspectives of container platform, security, and backend, and summarized what this shift means for actual service operations.

The event was held November 30–December 5, 2025 in Las Vegas, with 3,122 sessions running simultaneously across multiple hotels throughout the city. Reservations function more like priority admission than a guaranteed seat, and popular sessions give up their seats to Walk-up attendees if reservation holders don't arrive 10 minutes before start, with the practical tip being that moving between hotels safely requires budgeting over 1 hour.

On the technical side, the keynote featured notable announcements such as Graviton5, Trainium3 UltraServers, Nova 2, Nova Forge, Nova Act, and Amazon S3 Vectors. In particular, Nova Act is an Agent that automates browser-based UI workflows, enabling financial service scenarios where an Agent takes over lookup, copy, and paste tasks that people previously performed manually on internal systems.

From a platform perspective, the most impressive session was <AIM326 - Build and scale agentic apps with Amazon Bedrock>. The Policy feature of Bedrock AgentCore introduced there converts natural language policies into actual Cedar code to restrict the scope of an Agent's actions, while Evaluations monitors metrics such as Correctness, Helpfulness, Tool Selection Accuracy, and Safety in real time by integrating with CloudWatch.

This combination allows AI Agents to be treated as controllable systems rather than black boxes. For example, a refund Agent can be given a policy that only allows it to process amounts under $200, and if Tool Selection Accuracy drops after a model swap, it can be immediately detected as an anomaly and addressed.

On the container platform side, EKS Capabilities, introduced in <CON211 - What's new with Amazon EKS>, was presented as a turning point for easing operational burden. By elevating Argo CD, ACK, and KRO into AWS-managed features, repetitive tasks like patching, scaling, SSO integration, and CRD migration are reduced, and notably, ACK Resource Adoption allows existing Terraform resources to be migrated incrementally, which was an important point.

The security session <SEC408 - Securing Agentic AI> concretized the threats of the Agentic AI era. Based on OWASP threats such as Prompt Injection, Insecure Output Handling, and Excessive Agency, along with the MAESTRO 7-layer framework, least privilege, Human-in-the-loop, and continuous monitoring and logging were presented as core defense strategies.

Ultimately, the direction read from this event is clear. The full-scale rise of AI Agents, automation of platform management, and the security guardrails layered on top of them are the key pillars that will shape cloud strategy for the next year.