This project demonstrates how to implement different types of guardrails using the OpenAI Agent SDK with Trigger.dev for execution.

It serves as both:

• A practical guide for integrating the OpenAI Agent SDK with Trigger.dev for production AI workflows
• Educational examples of implementing three different types of guardrails for AI safety and control

Guardrails are safety mechanisms that run alongside AI agents to:

• Validate input before processing
• Check output before returning responses
• Monitor streaming content in real-time
• Prevent unwanted or harmful behavior

This project includes three different guardrail implementations:

Purpose: Validates user input before the agent processes it.

Example: A math tutor agent that only responds to mathematics-related questions. If you ask about anything else (like "What's the weather?"), the guardrail triggers and returns a refusal message.

How it works:

• Uses an agent-based guardrail to check if input is math-related
• Throws InputGuardrailTripwireTriggered exception when non-math topics are detected
• Returns a polite refusal instead of processing the request

Purpose: Validates the agent's response before returning it to the user.

Example: Ensures that a "Math Assistant" actually provides mathematical content in its responses. If the response doesn't contain sufficient math content, the guardrail triggers.

How it works:

• Agent generates a response first
• Guardrail agent evaluates if the response contains actual mathematical content
• Throws OutputGuardrailTripwireTriggered if response lacks math content
• Can either retry or return an error message

Purpose: Monitors content as it streams in real-time, allowing early termination.

Example: Checks if streaming responses use language too complex for a 10-year-old. If complex terms are detected while streaming, it immediately stops generation.

How it works:

• Streams response text to stdout in real-time
• Runs guardrail checks every N characters (configurable interval)
• Immediately stops streaming if guardrail triggers
• Provides detailed metrics about where/when the guardrail activated

1. Clone the repo and run npm install to install the dependencies
2. Create a virtual environment python -m venv venv
3. Activate the virtual environment:
   • On Mac/Linux: source venv/bin/activate
   • On Windows: venv\Scripts\activate
4. Install the Python dependencies pip install -r requirements.txt
5. Copy the project ref from your Trigger.dev dashboard and add it to the trigger.config.ts file
6. Run the Trigger.dev CLI dev command
7. Test the guardrail tasks in the dashboard
8. Deploy the tasks to production using the Trigger.dev CLI deploy command

• Trigger Tasks:

  • inputGuardrails.ts - Passes user prompts to Python script and handles InputGuardrailTripwireTriggered exceptions
  • outputGuardrails.ts - Runs agent generation and catches OutputGuardrailTripwireTriggered exceptions with detailed error info
  • streamingGuardrails.ts - Executes streaming Python script and parses JSON output containing guardrail metrics

• Python Implementations:

  • input-guardrails.py - Agent with @input_guardrail decorator that throws exceptions before main agent runs
  • output-guardrails.py - Agent with @output_guardrail decorator that validates generated responses using a separate guardrail agent
  • streaming-guardrails.py - Processes ResponseTextDeltaEvent streams with async guardrail checks at configurable intervals

• Configuration: trigger.config.ts - Uses the Trigger.dev Python extension

• OpenAI Agent SDK documentation
• OpenAI Agent SDK guardrails
• Trigger.dev documentation
• Trigger.dev Python extension