On-Site Report from re:Invent 2025

With hotels and convention centers across Las Vegas serving as venues, AWS re:Invent felt less like a simple conference and more like a technology event blanketing the entire city. During keynote hours, tens of thousands of engineers gathered in one space and reacted instantly to demos and messages, leaving a strong impression of scale that can only be felt in person.

The core message of re:Invent 2025 was clear. AWS is treating AI not as a "feature to adopt" but as a platform layered on top of a default premise. Beyond simple generative AI, Agentic AI—which handles planning, execution, and evaluation—took center stage, while cost optimization and operational stability continued to be emphasized.

One of the most impressive sessions was about making EKS operations Agentic. The flow was built to detect Slack notifications, search for past solutions, and, when needed, query cluster status in real time to resolve issues—going beyond the limitations of traditional RAG approaches by directly accessing current state via MCP. It was particularly practical to see the Managed EKS MCP server used to connect via IAM without having to spin up a separate server, and notable that the Strands Agents SDK was used to build an Orchestrator/Specialist structure with Slack messages as automatic triggers.

Cost efficiency was also factored in. Instead of processing every message heavily, Nova Micro first classified intent, and agents were only invoked when necessary. Organizational Tribal Knowledge scattered across teams was stored in S3 Vectors to make it reusable. On top of this, a separate Memory Agent collaborated with other agents via A2A, making it clear that the key to AI operations isn't a smarter model per se, but an architecture with real-time data access, accumulated knowledge, cost control, and well-separated roles.

The session on generative and agentic AI on Amazon EKS laid out a framework for viewing AI workloads from an operational perspective. As AI usage expands, Kubernetes—especially EKS—becomes an increasingly important foundation, for reasons including cost-to-performance optimization, portability across multi-cloud and on-premises environments, and the need to run existing services alongside inference and fine-tuning on a single platform.

The explanation that agent frameworks are ultimately just Python libraries—so packaging them into containers and running them on EKS isn't fundamentally different from existing services—was grounded and practical. Multiple tools and APIs can be bundled and reused as MCP servers, and GPU operations were explained through an analogy to seed, soil, and health management. The session also drew a distinction between Auto Mode and Karpenter, prompting reflection on how to balance configuration convenience against control.

The session on accelerating development and DevSecOps by combining Amazon Bedrock and Kiro showed the next stage of how AI enters development organizations. Feeding in a natural-language spec generates code, unit tests, documentation, and even architecture diagrams, and Bedrock AgentCore builds on this to carry the process through CI/CD, SAST/DAST/Dependency Check, code review, and vulnerability remediation.

What stood out most was the perspective of treating AI agents not as simple tools but as team members within the development team. Policy control limits the scope of their actions to manage cost, while Evaluations continuously measure accuracy and safety—making clear a direction where AI pushes beyond just speeding up development toward Shift-left Security and automated compliance.

Overall, this re:Invent wasn't so much about listing new technologies as it was about how to actually operate and integrate them within real systems. AI, cost, security, and operations were connected as a single flow rather than separate topics, and it prompted Banksalad to rethink, in moving toward being AI Native, which problems to apply AI to, how far to automate, and how to design cost and stability together.