openai-agents

jezweb's avatarfrom jezweb

Build AI applications with OpenAI Agents SDK - text agents, voice agents (realtime), multi-agent workflows with handoffs, tools with Zod schemas, input/output guardrails, structured outputs, and streaming. Deploy to Cloudflare Workers, Next.js, or React with human-in-the-loop patterns.Use when: building text-based agents with tools and Zod schemas, creating realtime voice agents with WebRTC/WebSocket, implementing multi-agent workflows with handoffs between specialists, setting up input/output guardrails for safety, requiring human approval for critical actions, streaming agent responses, deploying agents to Cloudflare Workers or Next.js, or troubleshooting Zod schema type errors, MCP tracing failures, infinite loops (MaxTurnsExceededError), tool call failures, schema mismatches, or voice agent handoff constraints.

⭐109starsšŸ”€16forksšŸ“View on GitHubšŸ•Updated Dec 28, 2025
nameopenai-agents
description|
Use whenbuilding agents with tools, voice agents with WebRTC, multi-agent workflows, or troubleshooting MaxTurnsExceededError, tool call failures, reasoning defaults, JSON output leaks.
user-invocabletrue

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)
    : await result.state.reject(result.interruption);
}

Example: human-in-the-loop-stream.ts


Realtime Voice Agents

Create:

import { RealtimeAgent } from '@openai/agents-realtime';
const voiceAgent = new RealtimeAgent({
  voice: 'alloy', // alloy, echo, fable, onyx, nova, shimmer
  model: 'gpt-5-realtime',
  tools: [weatherTool],
});

Browser Session:

import { RealtimeSession } from '@openai/agents-realtime';
const session = new RealtimeSession(voiceAgent, { apiKey: sessionApiKey, transport: 'webrtc' });
await session.connect();

CRITICAL: Never send OPENAI_API_KEY to browser! Generate ephemeral session tokens server-side.

Voice Handoffs: Voice/model must match across agents (cannot change during handoff)

Limitations:

  • Video streaming NOT supported: Despite camera examples, realtime video streaming is not natively supported. Model may not proactively speak based on video events. (Issue #694)

Templates:

  • templates/realtime-agents/realtime-agent-basic.ts
  • templates/realtime-agents/realtime-session-browser.tsx
  • templates/realtime-agents/realtime-handoffs.ts

References:

  • references/realtime-transports.md - WebRTC vs WebSocket

Framework Integration

Cloudflare Workers (experimental):

export default {
  async fetch(request: Request, env: Env) {
    // Disable tracing or use startTracingExportLoop()
    process.env.OTEL_SDK_DISABLED = 'true';

    process.env.OPENAI_API_KEY = env.OPENAI_API_KEY;
    const agent = new Agent({ name: 'Assistant', model: 'gpt-5-mini' });
    const result = await run(agent, (await request.json()).message);
    return Response.json({ response: result.finalOutput, tokens: result.usage.totalTokens });
  }
};

Limitations:

  • No voice agents
  • 30s CPU limit, 128MB memory
  • Tracing requires manual setup - set OTEL_SDK_DISABLED=true or call startTracingExportLoop() (Issue #16)

Next.js: app/api/agent/route.ts → POST handler with run(agent, message)

Templates: cloudflare-workers/, nextjs/


Error Handling (11+ Errors Prevented)

1. Zod Schema Type Errors

Error: Type errors with tool parameters.

Workaround: Define schemas inline.

// āŒ Can cause type errors
parameters: mySchema

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

Note: As of v0.4.1, invalid JSON in tool call arguments is handled gracefully (previously caused SyntaxError crashes). (PR #887)

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, // Increase limit
});

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

4. ToolCallError

Error: Tool execution fails.

Solution: Retry with exponential backoff:

for (let attempt = 1; attempt <= 3; attempt++) {
  try {
    return await run(agent, input);
  } catch (error) {
    if (error instanceof ToolCallError && attempt < 3) {
      await sleep(1000 * Math.pow(2, attempt - 1));
      continue;
    }
    throw error;
  }
}

5. Schema Mismatch

Error: Output doesn't match outputType.

Solution: Use stronger model or add validation instructions:

const agent = new Agent({
  model: 'gpt-5', // More reliable than gpt-5-mini
  instructions: 'CRITICAL: Return JSON matching schema exactly',
  outputType: mySchema,
});

6. Reasoning Effort Defaults Changed (v0.4.0)

Error: Unexpected reasoning behavior after upgrading to v0.4.0.

Why It Happens: Default reasoning effort for gpt-5.1/5.2 changed from "low" to "none" in v0.4.0.

Prevention: Explicitly set reasoning effort if you need it.

// v0.4.0+ - default is now "none"
const agent = new Agent({
  model: 'gpt-5.1',
  reasoning: { effort: 'low' }, // Explicitly set if needed: 'low', 'medium', 'high'
});

Source: Release v0.4.0 | PR #876

7. Reasoning Content Leaks into JSON Output

Error: response_reasoning field appears in structured output unexpectedly.

Why It Happens: Model endpoint issue (not SDK bug) when using outputType with reasoning models.

Workaround: Filter out response_reasoning from output.

const result = await run(agent, input);
const { response_reasoning, ...cleanOutput } = result.finalOutput;
return cleanOutput;

Source: Issue #844 Status: Model-side issue, coordinating with OpenAI teams

All Errors: See references/common-errors.md

Template: templates/shared/error-handling.ts


Orchestration Patterns

LLM-Based: Agent decides routing autonomously (adaptive, higher tokens) Code-Based: Explicit control flow with conditionals (predictable, lower cost) Parallel: Promise.all([run(agent1, text), run(agent2, text)]) (concurrent execution)


Debugging

process.env.DEBUG = '@openai/agents:*';  // Verbose logging
const result = await run(agent, input);
console.log(result.usage.totalTokens, result.history.length, result.currentAgent?.name);

āŒ Don't use when:

  • Simple OpenAI API calls (use openai-api skill instead)
  • Non-OpenAI models exclusively
  • Production voice at massive scale (consider LiveKit Agents)

Production Checklist

  • Set OPENAI_API_KEY as environment secret
  • Implement error handling for all agent calls
  • Add guardrails for safety-critical applications
  • Enable tracing for debugging
  • Set reasonable maxTurns to prevent runaway costs
  • Use gpt-5-mini where possible for cost efficiency
  • Implement rate limiting
  • Log token usage for cost monitoring
  • Test handoff flows thoroughly
  • Never expose API keys to browsers (use session tokens)

Token Efficiency

Estimated Savings: ~60%

Task Without Skill With Skill Savings
Multi-agent setup ~12k tokens ~5k tokens 58%
Voice agent ~10k tokens ~4k tokens 60%
Error debugging ~8k tokens ~3k tokens 63%
Average ~10k ~4k ~60%

Errors Prevented: 11 documented issues = 100% error prevention


Templates Index

Text Agents (8):

  1. agent-basic.ts - Simple agent with tools
  2. agent-handoffs.ts - Multi-agent triage
  3. agent-structured-output.ts - Zod schemas
  4. agent-streaming.ts - Real-time events
  5. agent-guardrails-input.ts - Input validation
  6. agent-guardrails-output.ts - Output filtering
  7. agent-human-approval.ts - HITL pattern
  8. agent-parallel.ts - Concurrent execution

Realtime Agents (3): 9. realtime-agent-basic.ts - Voice setup 10. realtime-session-browser.tsx - React client 11. realtime-handoffs.ts - Voice delegation

Framework Integration (4): 12. worker-text-agent.ts - Cloudflare Workers 13. worker-agent-hono.ts - Hono framework 14. api-agent-route.ts - Next.js API 15. api-realtime-route.ts - Next.js voice

Utilities (2): 16. error-handling.ts - Comprehensive errors 17. tracing-setup.ts - Debugging


References

  1. agent-patterns.md - Orchestration strategies
  2. common-errors.md - 9 errors with workarounds
  3. realtime-transports.md - WebRTC vs WebSocket
  4. cloudflare-integration.md - Workers limitations
  5. official-links.md - Documentation links

Official Resources


Version: SDK v0.4.1 Last Verified: 2026-01-21 Skill Author: Jeremy Dawes (Jezweb) Production Tested: Yes Changes: Added v0.4.0 breaking changes (Zod 4, reasoning defaults), invalid JSON handling (v0.4.1), reasoning output leaks, streaming HITL pattern, agent-as-tool context isolation, video limitations, Cloudflare tracing setup