议题方向（Track）

Track 1：Agent Infra — K8s 上的 Agent 运行时与基础设施

        Agent 的生产化部署本质上是 Kubernetes 上的一类新工作负载。如何安全隔离 Agent 代码执行、如何通过 CNCF 网关项目治理 MCP/A2A 流量、如何利用 K8s 原生的编排能力协调多 Agent 协作——本论坛聚焦 K8s 生态如何承载 Agent 工作负载。

Track 2：AI Observability & Reliability — CNCF 可观测生态在 AI/Agent 时代的演进   

        OpenTelemetry 是 CNCF 除 Kubernetes 之外最成功的 Graduated 项目。Agent 时代对可观测提出了全新挑战：GenAI 语义规范、多 Agent Trace 关联、推理性能指标采集。同时，混沌工程、SRE 实践也在向 AI/Agent 场景延伸。本论坛聚焦 CNCF 可观测与可靠性生态如何保障 AI/Agent 系统的质量。

Track 3：AI Infra & LLM Serving — K8s 上的推理引擎与 AI 基础设施

        大模型推理是 Agent 时代的算力底座，而 Kubernetes 是推理引擎的标准部署平台。vLLM、SGLang 等开源推理引擎如何在 K8s 上高效运行、GPU 调度如何优化、平台工程如何支撑 AI 工作负载——本论坛聚焦 K8s 作为 AI 推理基础设施的工程实践。

议题形式

• 标准演讲 (Presentation）：30 分钟单双人演讲
• 闪电演讲 (Lightning）：10 分钟演讲
  

议题结果将在征集结束后一到两周内通知。

一般情况下主办方不提供差旅费用。

如在议题提交时遇到问题，请联系邮箱(caiwei95@hotmail.com) 或者微信 (dagu139)。

---

Tracks

Track 1: Agent Infra — Agent Runtime & Infrastructure on Kubernetes

    The production deployment of Agents is essentially a new class of workload on Kubernetes. This track focuses on how the K8s ecosystem supports Agent workloads—from securely sandboxing Agent code execution, to governing MCP/A2A traffic through CNCF gateway projects, to leveraging native K8s orchestration capabilities for multi-agent collaboration.

Track 2: AI Observability & Reliability — The Evolution of the CNCF Observability Ecosystem in the AI/Agent Era

    OpenTelemetry is CNCF's most successful Graduated project after Kubernetes. The Agent era introduces entirely new observability challenges: GenAI semantic conventions, multi-agent trace correlation, and inference performance metric collection.

    At the same time, Chaos Engineering and SRE practices are extending into AI/Agent scenarios. This track focuses on how the CNCF observability and reliability ecosystem ensures the quality of AI/Agent systems.

Track 3: AI Infra & LLM Serving — Inference Engines & AI Infrastructure on Kubernetes

    LLM inference is the computing backbone of the Agent era, with Kubernetes serving as the standard deployment platform for inference engines.

    How open-source engines like vLLM and SGLang run efficiently on K8s, how GPU scheduling is optimized, and how platform engineering supports AI workloads—this track focuses on engineering practices for using K8s as AI inference infrastructure.

Session Formats

• Standard Presentation: 30 minutes, single or dual speakers
• Lightning Talk: 10 minutes, single speaker

Notification of results will be sent within one to two weeks after the submission deadline.

Please note that the organizers generally do not cover travel or accommodation expenses.

If you encounter any issues during the submission process, please reach out via email (caiwei95@hotmail.com) or WeChat (dagu139).