OpenAI Agents SDK

Build AI applications with text agents, voice agents (realtime), multi-agent workflows, tools, guardrails, and human-in-the-loop patterns.

Quick Start

npm install @openai/agents zod@4  # v0.4.0+ requires Zod 4 (breaking change)
npm install @openai/agents-realtime  # Voice agents
export OPENAI_API_KEY="your-key"

Breaking Change (v0.4.0): Zod 3 no longer supported. Upgrade to zod@4.

Runtimes: Node.js 22+, Deno, Bun, Cloudflare Workers (experimental)

Core Concepts

Agents: LLMs with instructions + tools

import { Agent } from '@openai/agents';
const agent = new Agent({ name: 'Assistant', tools: [myTool], model: 'gpt-5-mini' });

Tools: Functions with Zod schemas

import { tool } from '@openai/agents';
import { z } from 'zod';
const weatherTool = tool({
  name: 'get_weather',
  parameters: z.object({ city: z.string() }),
  execute: async ({ city }) => `Weather in ${city}: sunny`,
});

Handoffs: Multi-agent delegation

const triageAgent = Agent.create({ handoffs: [specialist1, specialist2] });

Guardrails: Input/output validation

const agent = new Agent({ inputGuardrails: [detector], outputGuardrails: [filter] });

Structured Outputs: Type-safe responses

const agent = new Agent({ outputType: z.object({ sentiment: z.enum(['positive', 'negative']) }) });

Text Agents

Basic: const result = await run(agent, 'What is 2+2?')

Streaming:

const stream = await run(agent, 'Tell me a story', { stream: true });
for await (const event of stream) {
  if (event.type === 'raw_model_stream_event') process.stdout.write(event.data?.choices?.[0]?.delta?.content || '');
}

Multi-Agent Handoffs

const billingAgent = new Agent({ name: 'Billing', handoffDescription: 'For billing questions', tools: [refundTool] });
const techAgent = new Agent({ name: 'Technical', handoffDescription: 'For tech issues', tools: [ticketTool] });
const triageAgent = Agent.create({ name: 'Triage', handoffs: [billingAgent, techAgent] });

Agent-as-Tool Context Isolation: When using agent.asTool(), sub-agents do NOT share parent conversation history (intentional design to simplify debugging).

Workaround: Pass context via tool parameters:

const helperTool = tool({
  name: 'use_helper',
  parameters: z.object({
    query: z.string(),
    context: z.string().optional(),
  }),
  execute: async ({ query, context }) => {
    return await run(subAgent, `${context}\n\n${query}`);
  },
});

Source: Issue #806

Guardrails

Input: Validate before processing

const guardrail: InputGuardrail = {
  execute: async ({ input }) => ({ tripwireTriggered: detectHomework(input) })
};
const agent = new Agent({ inputGuardrails: [guardrail] });

Output: Filter responses (PII detection, content safety)

Human-in-the-Loop

const refundTool = tool({ name: 'process_refund', requiresApproval: true, execute: async ({ amount }) => `Refunded $${amount}` });

let result = await runner.run(input);
while (result.interruption?.type === 'tool_approval') {
  result = await promptUser(result.interruption) ? result.state.approve(result.interruption) : result.state.reject(result.interruption);
}

Streaming HITL: When using stream: true with requiresApproval, must explicitly check interruptions:

const stream = await run(agent, input, { stream: true });
let result = await stream.finalResult();
while (result.interruption?.type === 'tool_approval') {
  const approved = await promptUser(result.interruption);
  result = approved
    ? await result.state.approve(result.interruption)