Inside LinkedIn's Kubernetes Ops Agent: Skills, Tools, and Guardrails

An AI agent that explains a failed deployment is useful. An agent that can help operate LinkedIn's Kubernetes-based compute platform spanning 500,000+ nodes and 5M+ pods needs more than a good prompt.

We built one operations agent for Slack, coding agent plugin, and automated deployment-failure workflows. Rather than build a separate MCP tool layer, we made every operational CLI used by our engineers friendly to agents. The same skills power every surface: they are loaded as coding-agent plugins, invoked for questions about applications or the compute fleet, used to auto-respond Slack support requests, and called by automated workflows to investigate deployment failures.

This talk explains how we separate platform knowledge, bounded actions, and multi-step operational workflows into composable skills. We use JIRA history and past support conversations to identify recurring operational problems and continuously improve those skills.

It also covers how we use server-side guardrails that make production use practical: rate limits, delete and scaled-down protection, access controls to prevent incorrect actions, and how peer approval keeps a human in the loop for required production changes.