How to Build AI Agents with the OpenAI Agents SDK

steps

1. Install the SDK using pip:

   bashCopy

   pip install openai-agents-sdk

2. Verify the installation with a basic Python import:

   pythonCopy

   import openai_agents_sdk
   print("SDK imported successfully!")

3. Set up your environment variables:

   • Create an .env file and include your OpenAI API key:
     Copy

     OPENAI_API_KEY=<your-api-key>

   • Use dotenv or any configuration library to load the key in your Python project.

steps

1. Import the necessary libraries and set up your API configuration:

   pythonCopy

   from openai_agents_sdk import AgentRunner, SimpleAgent
   
   agent = SimpleAgent(
       name="basic_agent",
       model="gpt-4",
       instructions="You are an assistant who writes haikus about programming."
   )

2. Write a prompt and execute the agent:

   pythonCopy

   def main():
       runner = AgentRunner()
       user_prompt = "Write a haiku about recursion."
       response = runner.run(agent, user_prompt)
       print(response)
   
   if __name__ == "__main__":
       main()

steps

1. Create a data class for structured outputs:

   pythonCopy

   from dataclasses import dataclass
   
   @dataclass
   class TravelPlan:
       destination: str
       duration: str
       budget: float
       activities: list
       notes: str

2. Configure the agent to use structured outputs:

   pythonCopy

   travel_agent = SimpleAgent(
       name="travel_assistant",
       model="gpt-4",
       instructions="Plan trips with a focus on activities, budget, and notes.",
       output_type=TravelPlan
   )

3. Test structured responses:

   pythonCopy

   def test_agent():
       runner = AgentRunner()
       query = "Plan a 5-day trip to Miami with a $2000 budget"
       result = runner.run(travel_agent, query)
       print(result)
   
   test_agent()

steps

1. Define a mock weather forecast tool:

   pythonCopy

   from openai_agents_sdk.tools import tool
   
   @tool
   def get_weather(city: str, date: str):
       """
       Fetch weather data for a city on a given date.
       """
       weather_data = {
           "Miami": "Sunny, 25-35°C",
           "Paris": "Cloudy, 12-22°C"
       }
       return weather_data.get(city, "Weather data unavailable.")

2. Integrate the tool into the agent:

   pythonCopy

   travel_agent_with_tools = SimpleAgent(
       name="travel_assistant_with_tools",
       model="gpt-4",
       instructions="Use weather data to suggest activities.",
       tools=[get_weather]
   )

3. Test the tool integration:

   pythonCopy

   runner = AgentRunner()
   query = "Suggest activities in Miami on 2023-12-01."
   response = runner.run(travel_agent_with_tools, query)
   print(response)

steps

1. Define specialized agents:

   pythonCopy

   flight_agent = SimpleAgent(
       name="flight_specialist",
       model="gpt-4",
       instructions="Provide flight options based on user preferences."
   )
   hotel_agent = SimpleAgent(
       name="hotel_specialist",
       model="gpt-4",
       instructions="Recommend hotels within user budget."
   )

2. Configure handoffs in the main travel agent:

   pythonCopy

   travel_agent = SimpleAgent(
       name="travel_assistant",
       model="gpt-4",
       instructions="Coordinate flights and hotels for user trips.",
       handoffs=[flight_agent, hotel_agent]
   )

3. Test inter-agent collaboration:

   pythonCopy

   queries = ["Find flights from NYC to London.", "Find Paris hotels under $300."]
   for q in queries:
       response = runner.run(travel_agent, q)
       print(response)

steps

1. Add input guardrails:

   pythonCopy

   @tool
   def budget_guardrail(destination: str, duration: int, budget: float):
       if budget < 500 or duration > 14:
           return f"Budget unrealistic for {destination} with a stay duration of {duration} days."
       return "Budget is feasible."

2. Create contextual user preferences:

   pythonCopy

   @dataclass
   class UserContext:
       preferred_airlines: list
       hotel_amenities: list
       max_budget: float

3. Apply guardrails and context:

   pythonCopy

   travel_agent_with_context = SimpleAgent(
       name="travel_assistant",
       model="gpt-4",
       instructions="Use context and guardrails to plan trips.",
       input_guardrails=[budget_guardrail],
       user_context=UserContext(preferred_airlines=["Delta"], hotel_amenities=["WiFi"], max_budget=2000)
   )