Chaos Engineering is a skill that gives an AI agent the structured methodology of a senior chaos engineer. Rather than producing generic best-practice checklists, it guides the agent through a four-phase procedure — Hypothesize, Perturb, Minimize, Learn — that treats infrastructure as a laboratory subject to controlled, evidence-based experiments.
For experiment design, the skill produces specific chaos experiments with measurable hypotheses, single-variable perturbations, and defined blast radii. This means instead of a vague recommendation like “add retries,” the agent will help design a concrete stress test to determine whether a system collapses under a retry storm.
For resilience auditing, the skill identifies hidden architectural failure modes including thundering herds, gray failures, and synchronized backoffs — patterns that standard reviews often miss.
For operational safety, it defines the human roles required to run experiments safely in production: Lead, Observer, and Abort Authority, along with specific readiness flags that must be satisfied before an experiment begins.
For post-incident work, it analyzes past incidents to produce “never again” experiments that verify a fix actually holds under realistic conditions.
The skill focuses on tail-risk scenarios at the P99 and P99.9 latency percentiles rather than averages, targeting the worst-case conditions that cause real outages. It is a reasoning and design skill; it does not execute experiments or interact with infrastructure directly.