The Metric Kubernetes Doesn't Give You — and Why It Kills AI Agents

When an AI agent runs out of memory in Kubernetes, the OOM killer doesn't care that it just spent 35 minutes building toward a result. The process dies, the context window is gone, and the work is unrecoverable. The problem is that cgroups don't expose the one metric that could have warned you: mem_available, the kernel's ground truth for how much memory your application actually has left. In this session, Jed Salazar (Head of Security, Edera) introduces kernel observability — the idea that if each container runs inside its own dedicated Linux kernel via a microVM, developers get access to every operating system metric for their specific application, not an aggregate view of the shared host. The session includes a live demo running Claude Code inside an Edera zone, simulating an AI agent filling its context window while a dashboard tracks the per-kernel metrics in real time. Topics covered: The OpenAI/Hugging Face incident: how an AI agent escaped its sandbox by finding zero-day vulnerabilities in an Artifactory instance during an Exploit Gym benchmark Why AI agents are not microservices: non-deterministic memory consumption, long-running stateful workloads, and irreplaceable context windows Why OOM kills are catastrophic for AI agents vs. recoverable for stateless replicas What mem_available actually measures and why cgroups don't track it The difference between free memory, cached memory, reclaimable memory, and anonymous memory — and why these distinctions determine whether your application gets OOM-killed PSI (Pressure Stall Information) as an early warning system for memory exhaustion Live demo: dedicated kernel metrics vs. shared kernel metrics under memory pressure Availability matrix: which operating system metrics exist in cgroups vs. a dedicated kernel This session is part of Edera's Sandbox Summer Webinar Series. 0:00 Introduction and OOM Kill Icebreaker 2:01 What Is Kernel Observability? 4:50 The OpenAI/Hugging Face AI Agent Incident 10:00 Edera: Untrusted Code Execution in MicroVMs 13:56 Demo Setup: Claude Code Hunting Zero-Days in the Go Standard Library 15:42 AI Agents Are Not Microservices 19:45 Live Demo: Dedicated Kernel Metrics vs. Shared Kernel Metrics 24:47 Mem Available: The Metric Cgroups Can't Give You 38:30 Memory Pressure and PSI as Early Warning Systems 45:23 How OOM Kills Actually Work 49:05 Availability Matrix: Cgroups vs. Dedicated Kernel