renocz

Agent Ops Guardrails — DevOps skill for Claude Code

DevOps community

Guardrails for an AI agent that runs real infrastructure: a multi-model review gate with a vote-computed verdict, and a Claude Code hook against accidental secret leaks.

How to install Agent Ops Guardrails

This entry records only its repository, not the path inside it, so there is no exact command to give. Open renocz/agent-ops-guardrails and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Agent Ops Guardrails does

Guardrails for an AI agent that runs real infrastructure: a multi-model review gate with a vote-computed verdict, and a Claude Code hook against accidental secret leaks. Tests, limits and field notes included.

Alternatives in DevOps

  • 16 Testing — Prompt 16: Add Test Infrastructure & Smoke Tests 2.3k ★
  • Cccc — Coordinate your coding agents like a group chat — read receipts, delivery tracking, and remote ops from your p 1.1k ★
  • WP CLI And Ops — WordPress CLI and operations management 902 ★

README

agent-ops-guardrails

[![tests](https://github.com/renocz/agent-ops-guardrails/actions/workflows/tests.yml/badge.svg)](https://github.com/renocz/agent-ops-guardrails/actions/workflows/tests.yml)

**An operating model for an AI agent that runs real infrastructure: the guardrails, the method, and the field notes.**

I let an AI agent run my homelab: a hypervisor, about forty containers, backups and a VPN. The agent is Claude Code; the reviewers in the council come from four different vendors. It has root, and it is good at the work. It is also confidently wrong a few times a week, and early on it printed secrets into its own transcript.

This repo collects the guardrails that grew around it. Each one exists because something went wrong once.

The principle in one picture

Each layer answers one way an AI ops agent fails. None of them is enough alone.

flowchart LR
    H(["👤 Human request"]) --> A["🤖 Agent analyses
(read-only)"] A --> P["📝 Written plan:
what, why, risk,
dry-run, rollback"] P -->|"decision that matters"| C{{"🏛️ council
4 models, 4 vendors"}} C -->|"FIX FIRST"| P C -->|"APPROVE
(+ isolated alerts)"| G{"👤 Human GO?"} P -->|"routine change"| G G -->|"no"| P G -->|"yes"| S["📸 snapshot gate
host takes a snapshot
(agent can't roll back)"] S --> E["⚙️ Execute + verify
with evidence"] E --> D["📚 Document
(change record)"] SG["🔐 secret-guard
every command and file read"] -.->|"watches"| A SG -.->|"watches"| E
Failure mode of the agent Guardrail Status
Confidently wrong plan: a wrong diagnosis, an inconsistent threshold, a missed risk **council/**: four models from different vendors review the written plan. The verdict is computed from their votes, and a lone blocking objection is never merged away. ✅ code
Prints a secret by accident (docker inspect, crontab -l, an error message carrying a URL with creden