A language model in a loop with tools, running until a goal is met or a limit is hit.
Observation → decision → action.
Agentic Cloud Ops
Operating cloud-native infrastructure where decisions are made by an agent per situation,
rather than written in advance in a pipeline or a manifest.
Tool / tool call
A typed function the model may call. In operations terms: kubectl get, a metrics query,
a kubectl apply.
Guardrail
A constraint that lives outside the model — RBAC, dry-run, approval gates, rate
limits. Prompts are guidance; guardrails are enforcement. A guardrail you can't verify is
enforced is decoration.
Human-in-the-loop
A required human approval before an action takes effect.
Runaway loop
The agent fixes what it just broke, forever. Countermeasure: a hard iteration ceiling.
Prompt injection
Instructions planted in data the agent reads — log messages, annotations, container names
— that change its behaviour.
Blast radius
How much can go wrong if the agent is wrong. Governed by RBAC and namespace scope.
MTTR
Mean Time To Recovery. The main argument for agents in operations.
Eval
A repeatable test of agent behaviour against known scenarios. Run before every prompt change.
MCP (Model Context Protocol)
An open protocol for exposing tools and data to models over a standard interface.
Toolset
A named group of MCP tools that can be enabled or disabled together, to reduce context
size and improve tool-selection accuracy.
k3d / K3s
Lightweight Kubernetes; k3d runs K3s in Docker. What we use in the workshop.
Talos
Immutable, API-driven Linux for Kubernetes. An alternative bootstrap, kept as an appendix.