Customer Support Agent

Full example with multiple tools, PII protection, and toxicity detection.

The source code for this example is available on GitHub: Customer Support Agent

This example demonstrates how to integrate the agent-control SDK into an existing application. It simulates a Customer Support Agent - a realistic enterprise scenario that shows the key patterns for protecting AI agents with server-defined controls.

Why This Example?

• Universally understood use case: Customer support is familiar to everyone
• Natural need for guardrails: PII protection, prompt injection defense
• Multiple operation types: LLM calls + tool calls (database, knowledge base, tickets)
• Enterprise-relevant: Shows patterns real companies would use

Quick Start

# 1. Install SDK and evaluators (first time only)
pip install -e sdks/python -e evaluators

# 2. Start all services (database, server, UI, demo controls)
./examples/customer_support_agent/demo.sh start

# 3. Run the demo
python examples/customer_support_agent/run_demo.py

The demo.sh start command:

• Starts PostgreSQL database
• Runs migrations
• Starts the API server (http://localhost:8000)
• Starts the UI (http://localhost:4000)
• Registers the agent with demo controls (PII detection, prompt injection)

When you open the UI, you’ll see the agent with controls already configured.

Other Commands

# Check status of all services
./examples/customer_support_agent/demo.sh status

# Stop all services
./examples/customer_support_agent/demo.sh stop

# Reset everything (deletes database, asks for confirmation)
./examples/customer_support_agent/demo.sh reset

Prerequisites

First-Time Setup

1. Install the SDK and evaluators:

   pip install -e sdks/python
   pip install -e evaluators
   

2. Install UI dependencies:

   cd ui
   pnpm install
   

Manual Setup (alternative to demo.sh)

If you prefer to run services manually:

1. Start the database (requires Docker)

   cd server
   docker compose up -d
   
2. Run database migrations

   cd server
   make alembic-upgrade
   
3. Start the server (Terminal 1)

   cd server
   make run
   

   Server runs at http://localhost:8000

4. Start the UI (Terminal 2)

   cd ui
   pnpm dev
   

   UI runs at http://localhost:4000

Running the Demo

After demo.sh start, the agent already has demo controls configured. Just run:

python examples/customer_support_agent/run_demo.py

Test different scenarios:

You: Hello, I need help with a refund
Agent: I understand you'd like a refund. Let me look into your order...

You: /test-pii
Running PII Detection Tests...
(Messages with SSN patterns will be blocked)

You: /test-injection
Running Prompt Injection Tests...
(Injection attempts will be blocked)

You: /quit
Goodbye!

Automated Mode

Run all test scenarios automatically:

python examples/customer_support_agent/run_demo.py --automated

Reset Agent Controls

To remove all controls from the agent (keeps the agent registered):

python examples/customer_support_agent/run_demo.py --reset

Adding Custom Controls

1. Open http://localhost:4000
2. Click on “Customer Support Agent” in the list
3. Click “Add Control” to create additional controls

See Example Controls below for configuration examples.

Available Commands

Command	Description
/help	Show all commands
/test-safe	Run safe message tests
/test-pii	Test PII detection (if control configured)
/test-injection	Test prompt injection detection
/lookup <query>	Look up customer (e.g., /lookup C001)
/search <query>	Search knowledge base
/ticket	Create a test support ticket
/quit	Exit the demo

Key Concepts

1. SDK Initialization

Initialize once at application startup:

import agent_control

agent_control.init(
    agent_name="Customer Support Agent",
    agent_id="646d5dea-c2e6-4453-b446-7035482b38e4",
    agent_description="AI-powered customer support assistant",
)

This:

• Registers the agent with the server
• Fetches the assigned policy and controls
• Enables the @control() decorator

2. Protecting Functions

Use the @control() decorator on any function you want to protect:

from agent_control import control

@control()
async def respond_to_customer(message: str) -> str:
    response = await llm.generate(message)
    return response

The decorator:

• Calls the server with check_stage="pre" before execution (validates input)
• Calls the server with check_stage="post" after execution (validates output)
• Raises ControlViolationError if a control triggers with “deny” action

3. Handling Violations

Catch ControlViolationError to provide graceful fallbacks:

from agent_control import ControlViolationError

try:
    response = await respond_to_customer(user_message)
except ControlViolationError as e:
    # Control triggered - return safe fallback
    print(f"Blocked by: {e.control_name}")
    response = "I cannot help with that request."

4. Controls are Server-Side

Important: Controls are defined on the server via the UI, not in code.

This design provides:

• Centralized management: Security team controls policies without code changes
• Instant updates: Change controls without redeploying agents
• Audit trail: Server logs all control evaluations
• Separation of concerns: Developers focus on features, security team on policies

Project Structure

customer_support_agent/
├── README.md                 # This file
├── demo.sh                   # Start/stop/reset script
├── setup_demo_controls.py    # Creates agent and demo controls
├── support_agent.py          # Main agent with SDK integration
└── run_demo.py               # Interactive demo runner

demo.sh

Manages the full demo lifecycle:

• start - Starts database, server, UI, and sets up demo controls
• stop - Stops all services
• reset - Deletes database and stops services
• status - Shows service status

setup_demo_controls.py

Creates the demo agent with pre-configured controls:

• block-ssn-in-output - Blocks responses containing SSN patterns
• block-prompt-injection - Blocks common injection attempts
• block-credit-card - Blocks credit card numbers in input

support_agent.py

Contains:

• SDK initialization
• Mock services (LLM, database, knowledge base, tickets)
• Protected functions with @control() decorator
• CustomerSupportAgent class with error handling

run_demo.py

Contains:

• Interactive chat loop
• Test command handlers (/test-pii, /test-injection, etc.)
• Automated test scenarios

Example Controls to Configure

The demo setup creates three controls automatically. Here are examples of additional controls you might add:

PII Detection (Post-check on output)

name: block-pii-in-output
scope:
  step_types: ["llm_inference"]
  stages: ["post"]
selector:
  path: output
evaluator:
  name: regex
  config:
    pattern: '\d{3}-\d{2}-\d{4}'  # SSN pattern
action:
  decision: deny
  message: "Response contains PII (SSN pattern)"

Prompt Injection (Pre-check on input)

name: block-prompt-injection
scope:
  step_types: ["llm_inference"]
  stages: ["pre"]
selector:
  path: input
evaluator:
  name: regex
  config:
    pattern: '(?i)(ignore.*instructions|system:|you are now)'
action:
  decision: deny
  message: "Potential prompt injection detected"

Toxic Content (Pre-check on input)

name: block-toxic-input
scope:
  step_types: ["llm_inference"]
  stages: ["pre"]
selector:
  path: input
evaluator:
  name: galileo.luna2
  config:
    stage_type: local
    metric: input_toxicity
    operator: gt
    target_value: 0.8
action:
  decision: deny
  message: "Inappropriate content detected"

Testing the Integration

1. Without controls: Run the demo without configuring any controls. All messages should pass through.

2. With PII control: Add a PII detection control, then run /test-pii. Messages with SSN patterns should be blocked.

3. With injection control: Add a prompt injection control, then run /test-injection. Injection attempts should be blocked.

Next Steps

• Explore the main examples for more integration patterns
• Read the SDK documentation

Overview

This example demonstrates a customer support agent protected by Agent Control with multiple controls:

• PII blocking — Prevents SSN and credit card numbers in responses
• Toxicity detection — Blocks toxic or harmful user messages via Luna-2
• Tool restriction — Limits which tools the agent can invoke

Source Code

See the full example on GitHub: Customer Support Agent