OpenAI Agents SDK

Integrate OpenAI's official Agents SDK with Agent Kernel.

Installation

pip install agentkernel[openai]

Basic Usage

from agents import Agent as OpenAIAgent
from agentkernel.cli import CLI
from agentkernel.openai import OpenAIModule

agent = OpenAIAgent(
    name="assistant",
    instructions="You are a helpful assistant.",
)

OpenAIModule([agent])

if __name__ == "__main__":
    CLI.main()

Multi-Agent System

from agents import Agent as OpenAIAgent
from agentkernel.openai import OpenAIModule

# Define agents with handoff capabilities
general_agent = OpenAIAgent(
    name="general",
    handoff_description="Agent for general questions",
    instructions="You provide general assistance.",
)

math_agent = OpenAIAgent(
    name="math",
    handoff_description="Specialist for math problems",
    instructions="You solve math problems.",
)

OpenAIModule([general_agent, math_agent])

Configuration

export OPENAI_API_KEY=sk-...
export OPENAI_MODEL=gpt-4  # Optional, override default

Tool Binding

Use OpenAIToolBuilder to bind plain Python functions as tools to your OpenAI agents:

from agents import Agent as OpenAIAgent
from agentkernel.openai import OpenAIModule, OpenAIToolBuilder

def get_weather(city: str) -> str:
    """Returns the weather for a given city."""
    return f"Weather in {city}: sunny, 25°C"

agent = OpenAIAgent(
    name="weather",
    instructions="You provide weather information.",
    tools=OpenAIToolBuilder.bind([get_weather]),
)

OpenAIModule([agent])

See Tools for the full guide on writing and binding tools.

Structured Output

Configure structured output with the OpenAI Agents SDK's output_type parameter. Agent Kernel detects the structured result and returns an AgentReplyAny whose content is the result as a dict; no re-parsing of text needed:

from agents import Agent as OpenAIAgent
from pydantic import BaseModel
from agentkernel.openai import OpenAIModule

class CalendarEvent(BaseModel):
    name: str
    date: str

agent = OpenAIAgent(
    name="extractor",
    instructions="Extract the calendar event from the text.",
    output_type=CalendarEvent,
)

OpenAIModule([agent])

Pydantic results are converted via model_dump(), and str(reply) returns the JSON-serialized content, so text-based consumers (chat integrations, logging) work unchanged. See Reply Types for how structured replies are surfaced, and Execution Hooks for how hooks receive them.

Streaming limitation

Structured output applies to non-streaming execution only. Streamed runs emit token-by-token text deltas.

Features

• ✅ Function calling
• ✅ Multi-agent handoff
• ✅ Streaming responses
• ✅ Structured output (output_type → AgentReplyAny)
• ✅ Session management
• ✅ Framework-agnostic tool binding

Example

See examples/cli/openai for complete examples.

For structured output, see examples/cli/openai_structured and examples/api/openai_structured (REST API + post-execution hook).