Official Resources

Python Documentation Python Source JavaScript Source Customer Service Demo Voice Sample Swarm Deprecation PyPI History NPM Package DataCamp Course

Key Features

  • Core Primitives: Agent, Runner, Session, Guardrail, and function_tool primitives for workflow orchestration and execution.
  • Structured Output: Auto-schema generation via Pydantic (Python) and Zod (JavaScript/TypeScript) for type-safe agent responses.
  • Streaming Support: Token-level streaming with stream=True (Python) or stream: true (JavaScript) for real-time interactions.
  • Voice & WebRTC: FastAPI sample (Python) and RealtimeAgent (JS/TS) for browser and Node.js voice applications.
  • Hosted Tools: WebSearch, FileSearch, CodeInterpreter tools available without OAuth setup.
  • Tracing & Observability: Built-in tracing exporters for Logfire, AgentOps, Braintrust, Keywords AI with plug-in architecture.
  • Session Management: SQLite, Redis, and custom session backends for conversation memory and state persistence.

Code Examples

Python Installation

bash 
pip install openai-agents==0.1.4
# Requires: openai>=1.35, pydantic>=2.7, typing_extensions>=4.10

JavaScript Installation

bash 
npm install @openai/agents@0.1.4 'zod@<=3.25.67'
# Requires: Node 22+, Deno 1.42+, Bun 1.1+

Simple Agent with Tracing (Python)

python 
import os, asyncio
from agents import Agent, Runner, SQLiteSession, trace

os.environ["OPENAI_API_KEY"] = "sk-..."

@trace
async def main():
    agent = Agent(
        name="Coder",
        instructions="Return concise Python.",
        tools=[lambda n: list(range(0, n, 2))]
    )
    session = SQLiteSession("demo")
    result = await Runner.run(agent, "Even numbers 0-10", session=session)
    print(result.final_output)  # [0, 2, 4, 6, 8, 10]

asyncio.run(main())

Agent Architecture (Python)

python 
# Core Primitives
Agent(name: str, instructions: str, tools: list[Tool]=[], 
      handoffs: list[Agent]=[], guardrails: list[Guardrail]=[], 
      output_type: Type=None)

Runner.run(agent, input, session=None, max_turns=10, stream=False)
SQLiteSession(session_id: str, db_path: str=":memory:")

@function_tool  # Auto-schema via Pydantic
@input_guardrail(func) / @output_guardrail(func)

Use Cases

  • Multi-agent workflows with tool support and memory
  • Customer service agents with handoff capabilities
  • Voice-enabled applications with WebRTC integration
  • Production agents with tracing and observability
  • Structured output applications with type safety

Pros & Cons

Advantages

  • Flexible model integration - OpenAI Responses API, Chat Completions, LiteLLM support
  • Rich tool & memory support - Built-in and custom tools with conversational memory
  • Production-ready features - Streaming, voice, tracing, guardrails
  • Multi-language support - Python and JavaScript/TypeScript SDKs
  • Active ecosystem - Coinbase AgentKit, AgentOps, DataCamp integrations

Disadvantages

  • API volatility - Breaking changes possible in 0.1.x versions
  • JavaScript/TypeScript maturity - TypeScript definitions still in preview
  • No visual designer - CLI-only, no drag-and-drop interface
  • Long-context limitations - Manual management via custom context required

Future Outlook & Integrations

  • API Stabilization [Q3 2025]: v0.2.0 release targeting stable APIs with breaking change protection
  • TypeScript Maturity [Q3 2025]: Full TypeScript definitions and improved JavaScript/TypeScript experience
  • Visual Designer [Not Public]: Drag-and-drop interface for agent workflow design
  • Long-Context Extensions [Future Extension]: Enhanced context management and memory handling capabilities