Seeing the Whole System: My Talk at AWS Community Day Shenzhen 2026

On October 17, I joined AWS Community Day Shenzhen 2026, “Bridge with Signal,” to share what we have been building in Apache SkyWalking. My talk was “Seeing the Whole System: Cloud-Native Infrastructure in 3D and AI Agents in Action.”

I started with a quick tour of what ten years of work by the SkyWalking community has made possible. We have built the ability to observe services and their dependencies across a complex system. The recent work on Horizon UI gives us new ways to see and explore that information.

I showed how the views fit together: a service topology for understanding calls, an instance map for examining who participates in a relationship, and a deployment view for looking inside a cluster such as BanyanDB. The 3D infrastructure map  brings applications, middleware, service mesh, and infrastructure into one scene. We can follow the same service through its application, Istio, and Kubernetes perspectives, with traffic, health, and alarms in context.

The aim is to make the system understandable at a glance, then let people inspect the details. I enjoyed showing how a modern visual interface can make the depth of SkyWalking’s observability easier to explore.

The main focus was Apache SkyWalking AI Sessionizer (ASZ) . AI agents choose their next steps as they work. A task can involve repeated model calls, tools, shell commands, and delegation to child agents. To understand the result, we need a way to revisit that work and the evidence it left behind.

We looked at replay for Claude Code, LangChain/LangGraph, and Kiro. ASZ assembles the available records into a conversation that we can navigate alongside its execution. We can inspect tool requests and results, follow delegated work, and examine the recorded context for model calls. Replay here means exploring the recorded session.

The Amazon Bedrock example made this concrete: an SRE agent investigates an incident, calls tools, and writes a report. With file-change recording enabled, we can connect the affected file and its diff to the writing tool, then read the agent’s response alongside the change. That connection is especially useful when we want to check what the agent actually accomplished.

I also shared a direction we are exploring: evaluating a whole conversation for task success, output quality, and whether the time and tokens were well spent. Capture and replay are available today. Conversation-level evaluation remains future work. As agents become our colleagues, I want us to understand both the outcomes and the work that produced them.

For anyone who wants to go deeper, my recent SkyWalking posts cover Claude Code replay , usage metrics and file changes , LangChain/LangGraph conversations , and recording sessions on Amazon Bedrock AgentCore . The demos and source are there to try, and I would love to hear what people find useful or still missing.

It was also great to meet friends and continue the conversations around the talk. We exchanged ideas about the latest progress in AI and coding, what it means for open source, and the progress and vision of the Apache Software Foundation. We also talked about Apache SkyWalking’s technical direction, including our ASZ integration with Kiro and Kiro Crew . We covered the session capture and replay work already in place for Kiro’s V3 agent engine, our plans to extend the integration, and the longer-term goal of evaluating whole conversations for outcomes, efficiency, and cost. Those conversations are a big part of why I enjoy meeting the community in person.

Thank you to the whole AWS China Group, and especially Yangling and Betty, for organizing and hosting the event. Thanks as well to everyone who joined my session and shared questions and ideas. I’m glad we had the chance to spend the day together in Shenzhen.