openai-agents

OpenAI Agents SDK Skill

Complete skill for building AI applications with OpenAI Agents SDK (JavaScript/TypeScript), covering text agents, realtime voice agents, multi-agent workflows, and production deployment patterns.

Quick Start

Installation

bun add @openai/agents zod@3
bun add @openai/agents-realtime  # For voice agents

Set environment variable:

export OPENAI_API_KEY="your-api-key"

Basic Text Agent

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

const agent = new Agent({
  name: 'Assistant',
  instructions: 'You are helpful.',
  tools: [tool({
    name: 'get_weather',
    parameters: z.object({ city: z.string() }),
    execute: async ({ city }) => `Weather in ${city}: sunny`,
  })],
  model: 'gpt-4o-mini',
});

const result = await run(agent, 'What is the weather in SF?');

Voice Agent & Multi-Agent

// Voice agent
const voiceAgent = new RealtimeAgent({
  voice: 'alloy',
  model: 'gpt-4o-realtime-preview',
});

// Browser session
const session = new RealtimeSession(voiceAgent, {
  apiKey: sessionApiKey, // From backend!
  transport: 'webrtc',
});

// Multi-agent handoffs
const triageAgent = Agent.create({
  handoffs: [billingAgent, techAgent],
});

17 Templates: templates/ directory has production-ready examples for all patterns.

Top 3 Critical Errors

1. Zod Schema Type Errors

Error: Type errors with tool parameters even when structurally compatible.

Workaround: Define schemas inline.

// ❌ Can cause type errors
parameters: mySchema

// ✅ Works reliably
parameters: z.object({ field: z.string() })

Source: GitHub #188

2. MCP Tracing Errors

Error: "No existing trace found" with MCP servers.

Workaround:

import { initializeTracing } from '@openai/agents/tracing';
await initializeTracing();

Source: GitHub #580

3. MaxTurnsExceededError

Error: Agent loops infinitely.

Solution: Increase maxTurns or improve instructions:

const result = await run(agent, input, {
  maxTurns: 20,
});

// Or improve instructions
instructions: `After using tools, provide a final answer.
Do not loop endlessly.`

All 9 Errors: Load references/common-errors.md for complete error catalog with workarounds.

When to Load References

Load reference files when working on specific aspects of agent development:

Agent Patterns (references/agent-patterns.md)

Load when:

• Designing multi-agent orchestration strategies
• Choosing between LLM-based vs code-based orchestration
• Implementing parallel agent execution
• Creating agents-as-tools patterns
• Need to understand when to use which orchestration pattern

Common Errors (references/common-errors.md)

Load when:

• Debugging agent issues beyond the top 3 errors above
• Implementing comprehensive error handling
• Encountering: GuardrailExecutionError, ToolCallError, Schema Mismatch, Ollama integration, webSearchTool failures, Agent Builder export bugs
• Building production error recovery patterns

Realtime Transports (references/realtime-transports.md)

Load when:

• Choosing between WebRTC vs WebSocket for voice agents
• Optimizing voice agent latency
• Debugging voice connection issues
• Understanding network/firewall requirements for voice
• Implementing custom audio sources/sinks

Cloudflare Integration (references/cloudflare-integration.md)

Load when:

• Deploying agents to Cloudflare Workers
• Understanding Workers limitations (CPU, memory, no voice)
• Implementing streaming in Workers
• Debugging Workers-specific issues
• Optimizing for Workers performance and costs

Official Links (references/official-links.md)

Load when:

• Need official documentation links
• Looking for examples or community resources
• Checking latest SDK versions
• Finding pricing information
• Need migration guides

Core Concepts Summary

Agents: LLMs equipped with instructions and tools.

Tools: Functions with Zod schemas that agents can call automatically.

Handoffs: Multi-agent delegation where agents route tasks to specialists.

Guardrails: Input/output validation for safety (content filtering, PII detection).

Structured Outputs: Type-safe responses using Zod schemas.

Streaming: Real-time event streaming for progressive responses.

Human-in-the-Loop: Require approval for specific tool executions (requiresApproval: true).

For detailed examples, see templates in templates/text-agents/ and templates/realtime-agents/.

Text Agents Quick Reference

// Basic
const result = await run(agent, 'Your question');

// Streaming
const stream = await run(agent, input, { stream: true });

// Structured output
const agent = new Agent({
  outputType: z.object({ sentiment: z.enum([...]), confidence: z.number() }),
});

Templates: templates/text-agents/ (8 templates)

Realtime Voice Agents Quick Reference

const voiceAgent = new RealtimeAgent({
  voice: 'alloy', // alloy, echo, fable, onyx, nova, shimmer
  model: 'gpt-4o-realtime-preview',
});

const session = new RealtimeSession(voiceAgent, {
  apiKey: sessionApiKey,
  transport: 'webrtc', // or 'websocket'
});

Voice handoff constraints: Cannot change voice/model during handoff.

Templates: templates/realtime-agents/ (3 templates) | Details: references/realtime-transports.md

Framework Integration Quick Reference

Cloudflare Workers (Experimental)

export default {
  async fetch(request: Request, env: Env) {
    const { message } = await request.json();
    process.env.OPENAI_API_KEY = env.OPENAI_API_KEY;

    const agent = new Agent({
      name: 'Assistant',
      instructions: 'Be helpful and concise',
      model: 'gpt-4o-mini',
    });

    const result = await run(agent, message, { maxTurns: 5 });

    return new Response(JSON.stringify({
      response: result.finalOutput,
      tokens: result.usage.totalTokens,
    }));
  },
};

Limitations: No realtime voice, CPU time limits (30s max), memory constraints (128MB).

Templates: templates/cloudflare-workers/ (2 templates)

Details: Load references/cloudflare-integration.md for complete Workers guide.

Next.js App Router

// app/api/agent/route.ts
import { NextRequest, NextResponse } from 'next/server';
import { Agent, run } from '@openai/agents';

export async function POST(request: NextRequest) {
  const { message } = await request.json();
  const agent = new Agent({ /* ... */ });
  const result = await run(agent, message);
  return NextResponse.json({ response: result.finalOutput });
}

Templates: templates/nextjs/ (2 templates)

Guardrails & Human-in-the-Loop

// Input/output guardrails
const agent = new Agent({
  inputGuardrails: [homeworkDetectorGuardrail],
  outputGuardrails: [piiFilterGuardrail],
});

// Human approval
const tool = tool({
  requiresApproval: true,
  execute: async ({ amount }) => `Refunded $${amount}`,
});

// Handle approval loop
while (result.interruption?.type === 'tool_approval') {
  result = (await promptUser(result.interruption))
    ? await result.state.approve(result.interruption)
    : await result.state.reject(result.interruption);
}

Templates: templates/text-agents/agent-guardrails-*.ts, agent-human-approval.ts

Orchestration Patterns Summary

LLM-Based: Agent decides routing autonomously. Use for adaptive workflows.

Code-Based: Explicit control flow. Use for predictable, deterministic workflows.

Parallel: Run multiple agents concurrently. Use for independent tasks.

Agents as Tools: Wrap agents as tools for manager LLM. Use for specialist delegation.

Details: Load `references/agent-patter