name: openai-agent-sdk-skill
description: Create OpenAI Agents SDK applications in TypeScript/JavaScript. Use when building AI agents, multi-agent systems, voice agents, or any agentic workflow with the OpenAI Agents SDK. Covers agents, tools, handoffs, guardrails, streaming, MCP, and realtime voice agents.

OpenAI Agents SDK Skill

Build powerful AI agents using the OpenAI Agents SDK for TypeScript/JavaScript.

Setup

# Create project
mkdir my-agent && cd my-agent
npm init -y

# Install dependencies
npm install @openai/agents zod@3

# Set API key
export OPENAI_API_KEY=sk-...

Or programmatically:

import { setDefaultOpenAIKey } from '@openai/agents';
setDefaultOpenAIKey(process.env.OPENAI_API_KEY!);

Basic Agent

IMPORTANT: Always use gpt-5-nano as the default model for most agents.

import { Agent, run } from '@openai/agents';

const agent = new Agent({
  name: 'Assistant',
  instructions: 'You are a helpful assistant.',
  model: 'gpt-5-nano', // Use gpt-5-nano for most agents
});

const result = await run(agent, 'Hello!');
console.log(result.finalOutput);

Agent Properties

Tools

Function Tools

import { Agent, tool, run } from '@openai/agents';
import { z } from 'zod';

const getWeather = tool({
  name: 'get_weather',
  description: 'Get weather for a city',
  parameters: z.object({
    city: z.string(),
  }),
  async execute({ city }) {
    return `Weather in ${city}: Sunny, 72°F`;
  },
});

const agent = new Agent({
  name: 'Weather Bot',
  instructions: 'Help with weather queries.',
  model: 'gpt-5-nano',
  tools: [getWeather],
});

const result = await run(agent, 'Weather in Tokyo?');

Tool with Context

import { Agent, tool, run, RunContext } from '@openai/agents';
import { z } from 'zod';

interface UserContext {
  userId: string;
  isAdmin: boolean;
}

const getUserData = tool({
  name: 'get_user_data',
  description: 'Get current user data',
  parameters: z.object({}),
  execute: async (_args, runContext?: RunContext<UserContext>) => {
    const ctx = runContext?.context;
    return `User: ${ctx?.userId}, Admin: ${ctx?.isAdmin}`;
  },
});

const agent = new Agent<UserContext>({
  name: 'User Bot',
  instructions: 'Help users with their data.',
  model: 'gpt-5-nano',
  tools: [getUserData],
});

const result = await run(agent, 'Show my data', {
  context: { userId: '123', isAdmin: true },
});

Hosted Tools

Available hosted tools that run on OpenAI servers:

import { Agent, webSearchTool, fileSearchTool, codeInterpreterTool } from '@openai/agents';

const agent = new Agent({
  name: 'Research Bot',
  instructions: 'Research and answer questions using web search.',
  model: 'gpt-5-nano',
  tools: [
    webSearchTool(),  // Internet search
    fileSearchTool('vs_your_vector_store_id'),  // Vector store search
    codeInterpreterTool(),  // Execute code
  ],
});

Structured Output

import { Agent, run } from '@openai/agents';
import { z } from 'zod';

const CalendarEvent = z.object({
  name: z.string(),
  date: z.string(),
  participants: z.array(z.string()),
});

const agent = new Agent({
  name: 'Calendar Extractor',
  instructions: 'Extract calendar events from text.',
  model: 'gpt-5-nano',
  outputType: CalendarEvent,
});

const result = await run(agent, 'Meeting with John tomorrow at 3pm');
console.log(result.finalOutput); // { name: '...', date: '...', participants: [...] }

Handoffs (Multi-Agent)

import { Agent, run } from '@openai/agents';

const billingAgent = new Agent({
  name: 'Billing Agent',
  instructions: 'Handle billing questions.',
  model: 'gpt-5-nano',
});

const supportAgent = new Agent({
  name: 'Support Agent',
  instructions: 'Handle technical support.',
  model: 'gpt-5-nano',
});

// Use Agent.create for proper type inference with handoffs
const triageAgent = Agent.create({
  name: 'Triage Agent',
  instructions: `Route users to the right agent:
    - Billing questions -> Billing Agent
    - Technical issues -> Support Agent`,
  model: 'gpt-5-nano',
  handoffs: [billingAgent, supportAgent],
});

const result = await run(triageAgent, 'I have a billing question');
console.log(result.finalOutput);
console.log(`Handled by: ${result.lastAgent.name}`);

Handoff with Input Filter

import { Agent, handoff } from '@openai/agents';
import { removeAllTools } from '@openai/agents-core/extensions';

const specialistAgent = new Agent({
  name: 'Specialist',
  instructions: 'Handle specialized queries.',
  model: 'gpt-5-nano',
});

const handoffObj = handoff(specialistAgent, {
  toolNameOverride: 'transfer_to_specialist',
  toolDescriptionOverride: 'Transfer to specialist for complex issues',
  inputFilter: removeAllTools, // Filter conversation history
});

Agents as Tools

Use an agent as a tool (keeps control with parent agent):

import { Agent, run } from '@openai/agents';

const summarizer = new Agent({
  name: 'Summarizer',
  instructions: 'Create concise summaries.',
  model: 'gpt-5-nano',
});

const summarizerTool = summarizer.asTool({
  toolName: 'summarize',
  toolDescription: 'Summarize provided text.',
});

const mainAgent = new Agent({
  name: 'Research Assistant',
  instructions: 'Help with research tasks.',
  model: 'gpt-5-nano',
  tools: [summarizerTool],
});

Guardrails

Input Guardrail

import { Agent, run, InputGuardrail, InputGuardrailTripwireTriggered } from '@openai/agents';
import { z } from 'zod';

const guardrailAgent = new Agent({
  name: 'Guardrail',
  instructions: 'Check if input is inappropriate.',
  model: 'gpt-5-nano',
  outputType: z.object({
    isInappropriate: z.boolean(),
    reason: z.string(),
  }),
});

const contentGuardrail: InputGuardrail = {
  name: 'Content Filter',
  execute: async ({ input, context }) => {
    const result = await run(guardrailAgent, input, { context });
    return {
      outputInfo: result.finalOutput,
      tripwireTriggered: result.finalOutput?.isInappropriate ?? false,
    };
  },
};

const agent = new Agent({
  name: 'Assistant',
  instructions: 'Help users.',
  model: 'gpt-5-nano',
  inputGuardrails: [contentGuardrail],
});

try {
  await run(agent, 'Some input');
} catch (e) {
  if (e instanceof InputGuardrailTripwireTriggered) {
    console.log('Guardrail tripped!');
  }
}

Streaming

import { Agent, run } from '@openai/agents';

const agent = new Agent({
  name: 'Storyteller',
  instructions: 'Tell engaging stories.',
  model: 'gpt-5-nano',
});

const result = await run(agent, 'Tell me a story', { stream: true });

// Stream text output
result.toTextStream({ compatibleWithNodeStreams: true }).pipe(process.stdout);

// Or listen to all events
for await (const event of result) {
  if (event.type === 'raw_model_stream_event') {
    console.log(event.data);
  }
}

await result.completed;

Human in the Loop (Tool Approval)

import { Agent, run, tool } from '@openai/agents';
import { z } from 'zod';
import readline from 'node:readline/promises';

const sensitiveTool = tool({
  name: 'delete_file',
  description: 'Delete a file',
  parameters: z.object({ filename: z.string() }),
  needsApproval: true, // Requires approval
  execute: async ({ filename }) => {
    return `Deleted ${filename}`;
  },
});

const agent = new Agent({
  name: 'File Manager',
  instructions: 'Manage files.',
  model: 'gpt-5-nano',
  tools: [sensitiveTool],
});

let result = await run(agent, 'Delete temp.txt');

while (result.interruptions?.length) {
  for (const interruption of result.interruptions) {
    const approved = await confirm(`Approve ${interruption.name}?`);
    if (approved) {
      result.state.approve(interruption);
    } else {
      result.state.reject(interruption);
    }
  }
  result = await run(agent, result.state);
}

Context Management

Dynamic Instructions

import { Agent, RunContext } from '@openai/agents';

interface UserContext {
  name: string;
  language: string;
}

function buildInstructions(ctx: RunContext<UserContext>) {
  return `User: ${ctx.context.name}. Respond in ${ctx.context.language}.`;
}

const agent = new Agent<UserContext>({
  name: 'Personalized Bot',
  instructions: buildInstructions,
  model: 'gpt-5-nano',
});

MCP (Model Context Protocol)

Stdio MCP Server

import { Agent, run, MCPServerStdio } from '@openai/agents';

const mcpServer = new MCPServerStdio({
  name: 'Filesystem',
  fullCommand: 'npx -y @modelcontextprotocol/server-filesystem ./files',
});

await mcpServer.connect();

const agent = new Agent({
  name: 'File Assistant',
  instructions: 'Help with files.',
  model: 'gpt-5-nano',
  mcpServers: [mcpServer],
});

try {
  const result = await run(agent, 'List files');
  console.log(result.finalOutput);
} finally {
  await mcpServer.close();
}

Hosted MCP (Remote)

import { Agent, hostedMcpTool } from '@openai/agents';

const agent = new Agent({
  name: 'Docs Assistant',
  instructions: 'Answer questions using docs.',
  model: 'gpt-5-nano',
  tools: [
    hostedMcpTool({
      serverLabel: 'docs',
      serverUrl: 'https://gitmcp.io/openai/codex',
    }),
  ],
});

Runner Configuration

import { Agent, Runner } from '@openai/agents';

const agent = new Agent({
  name: 'Assistant',
  instructions: 'Help users.',
  model: 'gpt-5-nano',
});

const runner = new Runner({
  model: 'gpt-5-nano', // Default model for all agents
  modelSettings: { temperature: 0.7 },
  workflowName: 'Customer Support',
  maxTurns: 15,
});

const result = await runner.run(agent, 'Hello');

Conversation History

import { Agent, run, user } from '@openai/agents';
import type { AgentInputItem } from '@openai/agents';

const agent = new Agent({
  name: 'Assistant',
  instructions: 'Be helpful.',
  model: 'gpt-5-nano',
});

let history: AgentInputItem[] = [];

async function chat(message: string) {
  history.push(user(message));
  const result = await run(agent, history);
  history = result.history;
  return result.finalOutput;
}

await chat('My name is Alice');
await chat('What is my name?'); // Will remember Alice

Lifecycle Hooks

import { Agent } from '@openai/agents';

const agent = new Agent({
  name: 'Verbose Agent',
  instructions: 'Be helpful.',
  model: 'gpt-5-nano',
});

agent.on('agent_start', (ctx, agent) => {
  console.log(`[${agent.name}] started`);
});

agent.on('agent_end', (ctx, output) => {
  console.log(`Agent produced: ${output}`);
});

Voice/Realtime Agents

Basic Voice Agent

import { RealtimeAgent, RealtimeSession } from '@openai/agents/realtime';

const agent = new RealtimeAgent({
  name: 'Voice Assistant',
  instructions: 'Help users with voice queries.',
});

const session = new RealtimeSession(agent, {
  model: 'gpt-realtime',
});

// Connect with ephemeral key (browser/WebRTC)
await session.connect({
  apiKey: 'ek_...', // Ephemeral key from backend
});

// Or with API key (server/WebSocket)
await session.connect({
  apiKey: process.env.OPENAI_API_KEY!,
});

Voice Agent with Tools

import { RealtimeAgent, RealtimeSession, tool } from '@openai/agents/realtime';
import { z } from 'zod';

const getWeather = tool({
  name: 'get_weather',
  description: 'Get weather for a city',
  parameters: z.object({ city: z.string() }),
  execute: async ({ city }) => `Weather in ${city}: Sunny`,
});

const agent = new RealtimeAgent({
  name: 'Weather Voice Bot',
  instructions: 'Answer weather questions.',
  tools: [getWeather],
});

const session = new RealtimeSession(agent, { model: 'gpt-realtime' });

Voice Agent Events

session.on('audio', (event) => {
  // Handle audio output (WebSocket only)
});

session.on('audio_interrupted', () => {
  // User interrupted the agent
});

session.on('history_updated', (history) => {
  // Conversation history changed
});

session.on('tool_approval_requested', (ctx, agent, request) => {
  session.approve(request.approvalItem);
  // or session.reject(request.rawItem);
});

Complete Multi-Agent Example

import { Agent, run, tool } from '@openai/agents';
import { z } from 'zod';

// Tools
const searchOrders = tool({
  name: 'search_orders',
  description: 'Search customer orders',
  parameters: z.object({ query: z.string() }),
  execute: async ({ query }) => `Found orders matching: ${query}`,
});

const processRefund = tool({
  name: 'process_refund',
  description: 'Process a refund',
  parameters: z.object({ orderId: z.string(), reason: z.string() }),
  execute: async ({ orderId, reason }) => `Refund processed for ${orderId}`,
});

// Specialist agents
const orderAgent = new Agent({
  name: 'Order Specialist',
  instructions: 'Handle order inquiries and searches.',
  model: 'gpt-5-nano',
  tools: [searchOrders],
});

const refundAgent = new Agent({
  name: 'Refund Specialist',
  instructions: 'Handle refund requests professionally.',
  model: 'gpt-5-nano',
  tools: [processRefund],
});

// Triage agent with handoffs
const triageAgent = Agent.create({
  name: 'Customer Support',
  instructions: `You are the first point of contact for customers.
    - Order questions -> Order Specialist
    - Refund requests -> Refund Specialist
    - General questions -> Answer directly`,
  model: 'gpt-5-nano',
  handoffs: [orderAgent, refundAgent],
});

// Run the multi-agent system
async function main() {
  const result = await run(triageAgent, 'I want a refund for order #12345');
  console.log(`Response: ${result.finalOutput}`);
  console.log(`Handled by: ${result.lastAgent.name}`);
}

main();

Best Practices

1. Use gpt-5-nano for most agents - it's fast and cost-effective
2. Keep agents focused - one responsibility per agent
3. Use handoffs for multi-domain tasks
4. Add guardrails for production safety
5. Enable streaming for better UX
6. Use structured output when you need typed responses
7. Implement human-in-the-loop for sensitive operations

Error Handling

import {
  run,
  MaxTurnsExceededError,
  InputGuardrailTripwireTriggered,
  ToolCallError
} from '@openai/agents';

try {
  const result = await run(agent, input);
} catch (e) {
  if (e instanceof MaxTurnsExceededError) {
    console.log('Too many turns');
  } else if (e instanceof InputGuardrailTripwireTriggered) {
    console.log('Guardrail blocked input');
  } else if (e instanceof ToolCallError) {
    console.log('Tool execution failed');
  }
}