Below, continues my reflections session by session — what I learnt, what resonated, and why I think it matters for practitioners in New Zealand and beyond.
Secure MCP on AWS
Speaker: Kobe Xu - Cloud Solutions Architect @ Deloitte - AWS Cloud Club NZ Lead
This was one of my favourite deep-dive sessions. Kobe unpacked Model Context Protocol (MCP), developed by Anthropic in late 2024.
The analogy was neat: MCP is like a USB-C port for AI applications — a standard way to connect AI to external systems. By early 2025, there were already over 1,000 community-built MCP servers.
He broke down the architecture:
- MCP host (the AI application with the MCP client).
- MCP servers (resource providers with primitives like tools, resources, prompts).
- Evolving separation between MCP servers and authorisation servers.
Security risks loomed large: misconfigured authentication, tool poisoning, prompt injection, and the lack of standard enterprise authentication flows.
Solutions included:
- OAuth 2.1 integration with Cognito.
- Fine-grained RBAC and least privilege.
- Zero-trust principles.
- Logging, monitoring, and hardened servers.
- Enterprise-level governance with MCP Gateways.
The enterprise use cases resonated: AI agent governance, dynamic tool discovery, unified access control.
Takeaway for technologists: MCP is set to be the plumbing layer of agentic AI applications, but without strong governance it could just as easily become the attack surface.
Accelerating Application Development and Cluster Management with the Amazon EKS MCP Server
Speaker: Carlo Columna - Senior Engineer @ Xero AWS Community Builder
Carlo’s session was a good example of how quickly the Model Context Protocol (MCP) is moving from being an interesting new standard into a practical enabler of real engineering workflows. His framing was clear: MCP is not just another AI gimmick. It is a game-changer because it reduces hallucinations, enables true automation, solves integration nightmares, and provides security and control in one standardised model.
The bulk of his talk was about the EKS MCP server and how it fundamentally changes the way developers can interact with Kubernetes. Instead of juggling CLI tools, remembering kubectl syntax, or manually applying YAML, Carlo demonstrated that these routine but critical tasks can now be abstracted into MCP server tools.
What impressed me most was the live demo. Carlo walked through a full workflow, executed via spec-driven, agentic development using Amazon Q and EKS MCP:
Each of these steps would normally require context-switching, manual scripting, and repetitive CLI commands. But with the MCP server, they became structured tasks in an agentic workflow.
The session underlined a bigger theme: Kubernetes is not going away, but how we operate and manage it is evolving. Developers are no longer restricted to point-and-click dashboards or opaque YAML — with MCP, they gain a standardised interface to drive cluster operations safely, with AI as a helper.
My reflection: this talk gave me confidence that MCP isn’t just theoretical. It is already enabling workflows that make Kubernetes management faster, more consistent, and easier to automate — without taking away the precision and control engineers need. For teams running Kubernetes at scale, this could be the beginning of a new operating model where “describe what you want” replaces “type the right command.”
The World of Agentic AI: A Developer’s Guide
Speaker: Donnie Prakoso - Principal Developer Advocate @ AWS Web Services
Donnie’s talk was one of the richest of the day because it attempted to map out the broader landscape of agentic AI and what it means for developers. He began by distinguishing between two paths of AI-driven development currently emerging:
- AI-Managed Development: where AI is largely autonomous, building and maintaining software with minimal human oversight.
- AI-Assisted Development: where humans remain firmly in control, using AI for targeted tasks while retaining intellectual ownership.
Most of us in the room were naturally leaning toward AI-assisted approaches, but Donnie made it clear that both paradigms will coexist.
To give structure to this new world, he introduced the AI-Driven Development Lifecycle (AI-DLC). This was useful because it mapped AI involvement across the entire software development lifecycle:
- Inception/Design phase: build context on existing code, expand intent with user stories, plan with work units.
- Development phase: generate domain models, add architecture, produce code and tests, generate IaC templates.
- Operations phase: deploy environments with IaC, manage incidents, support BAU.
In this framing, AI acts as orchestrator and co-pilot, but the developer remains the ultimate decision-maker.
Donnie also introduced us to AI-first IDEs such as Cursor and KIRO (AWS’s own experimental IDE). What caught my attention with KIRO was its philosophy:
- Spec-based development: everything starts with structured requirements.
- Steering: developers can let the agent “just do it” but then intervene, pausing or redirecting when needed.
- Automation with agent hooks: giving developers automation without sacrificing oversight.
The conversation naturally circled back to MCP, which he positioned as the connective tissue for agentic workflows — allowing AI models to interact with tools, resources, and data sources in a standard way. He also spoke of agentic communication — how MCP, stand agents, and microservices could form the building blocks of distributed systems.
A particularly exciting preview was Amazon Bedrock AgentCore, currently in limited preview. AgentCore will provide:
- A runtime for deploying agents securely at scale.
- Tooling and memory support via gateways and interpreters.
- Observability features to monitor agent activity.
This hinted at a future where enterprises could deploy agents with the same governance and scalability they expect from microservices today.
Donnie wrapped up with some grounded reminders: agentic AI doesn’t absolve us of distributed systems fundamentals. We still need:
- Usage plans to throttle and gate requests.
- Idempotency to avoid duplicate operations.
- Deterministic inputs to manage non-deterministic outputs from AI.
My reflection: Donnie’s session painted the most comprehensive picture of where we’re heading. Agentic AI is not just about building apps faster — it is about reshaping software architecture itself. It introduces new abstractions but doesn’t replace the principles we already know. For me, the key insight was that agentic workflows are the next evolutionary layer, sitting on top of our distributed systems knowledge, not replacing it. The challenge is ensuring that as we adopt these abstractions, we don’t forget the fundamentals underneath.
Networking and Community Spirit
No Community Day is complete without the hallway track. I connected with AWS Heroes, Ambassadors, and fellow community members who have been shaping these events. These conversations reminded me why this ecosystem thrives: it’s not just about tech, but about peer-to-peer exchange of lived experiences.
For me personally, this year highlighted how agentic AI has matured significantly. The playful “vibing” phase of conversational coding won’t scale. Instead, spec-driven, workflow-based development is here to stay — at least until the next paradigm shift.
Final Reflections
Walking away from the event, I felt both energised and grounded. Energised because the pace of innovation is accelerating; grounded because the themes that matter most — documentation, governance, orchestration — remain timeless. For those working in this space in New Zealand and beyond, the next 12–18 months will be crucial for putting these concepts into practice.
A sincere thank you to the organizing committee and all the speakers for successfully orchestrating the remarkable third edition of Community Day Aotearoa.
Resources
AWS Community Day NZ - 2025
Speaker slides (to be updated once available)
If you missed Part 1 of this article, it can be found here.
#AWSCDNZ25 #AWSCommunity #awscommunitybuilders #Wellington #NewZealand #EventHighlights
Comments
Post a Comment