Building Production Multi-Agent Workflows with the OpenAI Agents SDK in TypeScript (2026 Guide)

3. Multi-Agent Orchestration Patterns

There are two primary patterns. Choose deliberately.

3.1 Handoffs (specialist takes over)

Use handoffs when the specialist should own the rest of the conversation and speak directly to the user.

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

const historyTutor = new Agent({
  name: 'History Tutor',
  instructions: 'Explain historical events clearly with context and dates.',
});

const mathTutor = new Agent({
  name: 'Math Tutor',
  instructions: 'Solve math problems step by step and show reasoning.',
});

const triage = Agent.create({
  name: 'Triage',
  instructions: 'Route the user question to the correct specialist.',
  handoffs: [historyTutor, mathTutor],
});

const result = await run(triage, 'When did the Roman Empire fall?');
console.log(result.finalOutput);
console.log('Handled by:', result.lastAgent?.name);

Agent.create keeps TypeScript types aligned across handoff graphs. The runner automatically transfers conversation context to the chosen specialist.

3.2 Agents as tools (manager stays in control)

Use this when you want one agent to own the final answer and combine results from specialists without giving them the user-facing conversation.

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

const researcher = new Agent({
  name: 'Researcher',
  instructions: 'Find concise, factual answers. Prefer primary sources.',
});

const writer = new Agent({
  name: 'Writer',
  instructions: 'Turn research notes into clear technical prose.',
});

const manager = new Agent({
  name: 'Manager',
  instructions: 'Coordinate research and writing. Always return a polished final answer.',
  tools: [
    researcher.asTool({
      toolName: 'research',
      toolDescription: 'Research a technical topic and return key facts',
    }),
    writer.asTool({
      toolName: 'write',
      toolDescription: 'Turn research notes into a short technical article section',
    }),
  ],
});

const result = await run(
  manager,
  'Write a short section explaining why multi-agent systems improve reliability over a single general agent.'
);
console.log(result.finalOutput);

You can mix both patterns: a triage agent hands off to a specialist, and that specialist still uses other agents as tools for bounded subtasks.