1 · What is Agentic Cloud Ops?¶
12 minutes · presentation and discussion
The concept¶
What separates an agent from traditional automation (CI/CD) and GitOps?
A pipeline executes a predefined sequence. GitOps converges towards a declared desired state. An agent, by contrast, runs a loop:
flowchart LR
O[Observation] --> D[Decision]
D --> A[Action]
A --> O
The difference is that the decision isn't written in advance. You give it a goal, not a procedure.
| CI/CD | GitOps | Agentic Ops | |
|---|---|---|---|
| Input | A trigger | Desired state in Git | A goal in natural language |
| Logic | Written up front | Declarative diff | Derived per run |
| Determinism | High | High | Low |
| Handles the unexpected | No | No | Often |
| Auditability | Pipeline log | Git history | Requires deliberate work |
The opportunities¶
- Lower MTTR — the agent correlates events, logs and manifests in seconds.
- Automatic optimisation — resource limits, autoscaling, cleanup.
- Levelling knowledge — juniors get a senior's debugging patterns on tap.
The dangers¶
These are the real failure modes
- Hallucinations that delete namespaces or CRDs.
- No determinism — same prompt, different result.
- Runaway loops — the agent fixes what it just broke, forever.
- Prompt injection via log messages, annotations or container names.
- Blast radius — one over-privileged service account is enough.
The guardrail that wasn't¶
A real example from this year, and the single best argument for everything that follows.
CVE-2026-46519 — mcp-server-kubernetes, CVSS 8.8
Three environment variables (ALLOW_ONLY_READONLY_TOOLS,
ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS, ALLOWED_TOOLS) filtered which tools appeared in
the tools/list response.
But the filter was never enforced at execution time. A client could skip discovery
and call kubectl_delete directly over the MCP protocol.
Fixed in 3.6.0 — the same day it was reported.
The server said it was read-only. The tool list looked read-only. The cluster was not.
A guardrail you can't verify is enforced is decoration.
Keep this in mind during Lab 2, where we build guardrails of our own.
Success criteria¶
- Human-in-the-loop at first. Approve every write until you know the agent's patterns.
- Good RBAC. Don't hand out
cluster-adminuncritically. Start read-only, widen deliberately. - Genuinely good system prompts. Define scope, forbidden operations, reporting format.
- Log everything. Every tool call, every argument, every result.
- A rollback that doesn't depend on the agent.
Discussion¶
Where in your own stack would you let an agent loose today — and what would stop you?