An AI agent recommended the same fix four times.
Confident. Well-reasoned. Impossible.
Letting AI operate production Azure: what it gets wrong
Damian Flynn · Microsoft MVP, Azure · Innofactor
Damian Flynn
Practice Lead & Principal Consultant (Azure), Innofactor
30+ years in enterprise IT · cloud since 2009 · Mayo, Ireland
Azure Infrastructure as Code · Azure Networking · 17 years an MVP
AI doesn't fail loudly.
It fails plausibly.
Six incidents. Same steps each time.
The story
What happened in production. Customers anonymised.
The trap
The agent's mistake, replayed from a real session.
The guardrail
A real check, run live, catching it.
The rule
One line you can take home.
You launch Claude and fire away. Where does the guardrail fire?
Not a Claude trick. Every harness has a Layer B.
| Harness | Where it lives | Event | How it blocks |
|---|---|---|---|
| Claude Code | .claude/settings.json | PreToolUse | exit 2 |
| GitHub Copilot CLI | .github/hooks/*.json | preToolUse | "permissionDecision": "deny" |
| OpenAI Codex | .codex/hooks.json | PreToolUse | deny |
| Gemini CLI | .gemini/settings.json | BeforeTool | exit 2 |
| Pi | .pi/extensions/*.ts | pi.on("tool_call") | { block: true } |
| OpenCode | .opencode/plugins/*.ts | tool.execute.before | throw |
Same guardrail script, six doors. Seen blocking live this week: 5 of 6. Untrusted folder? The hook silently doesn't fire.
The six
Empty is not unset
Green is not working
The world moved while I was thinking
Telephone
Who reviews the reviewer?
A freeze is a promise, not a lock
Empty is not unset
Settable: never.
Where the agent looked, and where the setting really lives
OpenVPN
Green is not working
Receivers: 0.
Three green lights, three dead systems
≠ capability
The world moved while I was thinking
~15 minutes.
An operator was mid-change. The agent was reading an old map.
Can my team share this? Yes. It's just files.
Your machine
~/.claude/ · settings.local.json
Your experiments. Nobody else gets them, and nobody reviews them.
The repo
AGENTS.md · .claude/settings.json · hooks/ · skills/
Committed and reviewed in a PR, like code. Everyone who clones gets the same guardrails.
Managed policy
managed settings · requirements.toml · a shared plugin
Pushed by MDM or a template repo. Users can't switch it off.
Context travels the same way: conventions in AGENTS.md, runbooks and incident notes in the repo. Secrets never.
Telephone
Three tickets.
Every check was right. The retelling wasn't.
from memory
Who reviews the reviewer?
Actually: by design.
A second model checks the first. Who checks the second?
A freeze is a promise, not a lock
79 subscriptions.
Two doors into one firewall. Only one was supposed to be open.
Three of six were broken open by one plain question
“Don't the clients set DNS in their local profiles?”
“Can anyone still update it with the old code?”
And a simple question about whether a hosting option even had an address.
The guardrails exist to ask those questions every time, not when someone happens to.
Take these home
AI doesn't fail loudly. So check quietly, every time.
damianflynn.com · github.com/damianflynn · linkedin.com/in/damianflynn