Example: Network Automation Agent

This example shows how to use AgenticGuard with an AI agent that manages network infrastructure — a high-risk scenario where every action must be audited and destructive operations must be blocked.

Scenario

An LLM-powered network agent that can:

• Query device status (read-only, safe)
• Update firewall rules (write, needs audit)
• Restart services (destructive, needs approval)

Policy

version: "1"
rules:
  # Block all destructive actions
  - tool: network
    action: restart
    decision: deny
    message: "Service restarts require manual approval"

  - tool: network
    action: shutdown
    decision: deny

  # Require reasoning for any write
  - tool: network
    action: update
    require_reasoning: true
    message: "Firewall rule updates require agent reasoning"

  - tool: network
    action: create
    require_reasoning: true

  # Allow reads freely
  - tool: network
    action: get
    decision: allow

  - tool: network
    action: list
    decision: allow

Agent code

from agentic_guard import AgenticGuard, PolicyViolationError

gate = AgenticGuard(
    api_key="ag_xxxxxxxxxxxx",
    base_url="https://api.agentic-guard.com",
    mode="enforce",
)

def query_device_status(device_id: str) -> dict:
    return gate.call(
        tool="network",
        action="get",
        tool_url=f"https://network-api.internal/devices/{device_id}",
        method="GET",
        task_id="status_check",
    )

def update_firewall_rule(rule_id: str, new_config: dict, reasoning: str) -> dict:
    return gate.call(
        tool="network",
        action="update",
        tool_url=f"https://network-api.internal/firewall/{rule_id}",
        method="PUT",
        body=new_config,
        reasoning=reasoning,
        task_id="firewall_update",
    )

def restart_service(service_name: str) -> dict:
    try:
        return gate.call(
            tool="network",
            action="restart",
            tool_url=f"https://network-api.internal/services/{service_name}/restart",
            method="POST",
        )
    except PolicyViolationError as e:
        # Signal the LLM to request manual approval
        return {
            "error": "requires_approval",
            "message": e.message,
            "action": "restart",
            "target": service_name,
        }

What the audit log shows

timestamp           tool      action   decision  reasoning
2026-07-16 10:01    network   get      allowed   —
2026-07-16 10:02    network   update   allowed   "Blocking port 8080 per security scan #42"
2026-07-16 10:03    network   restart  denied    —   ← blocked, manual review triggered

Key takeaways

• Read operations are allowed freely (low risk)
• Write operations require reasoning (creates an audit trail)
• Destructive operations are blocked entirely — the agent returns a structured signal for human approval