[OpenAI DevDay 2025] AgentKitのAgent BuilderとChatKitを使ってみた

import { z } from "zod";
import { Agent, AgentInputItem, Runner } from "@openai/agents";

const WebResearchAgentSchema = z.object({ companies: z.array(z.object({ company_name: z.string(), industry: z.string(), headquarters_location: z.string(), company_size: z.string(), website: z.string(), description: z.string(), founded_year: z.number() })) });
const SummarizeAndDisplaySchema = z.object({ company_name: z.string(), industry: z.string(), headquarters_location: z.string(), company_size: z.string(), website: z.string(), description: z.string(), founded_year: z.number() });
const webResearchAgent = new Agent({
  name: "Web research agent",
  instructions: "You are a helpful assistant. Use web search to find information about the following company I can use in marketing asset based on the underlying topic.",
  model: "gpt-5-mini",
  outputType: WebResearchAgentSchema,
  modelSettings: {
    reasoning: {
      effort: "low"
    },
    store: true
  }
});

const summarizeAndDisplay = new Agent({
  name: "Summarize and display",
  instructions: `Put the research together in a nice display using the output format described.
`,
  model: "gpt-5",
  outputType: SummarizeAndDisplaySchema,
  modelSettings: {
    reasoning: {
      effort: "minimal"
    },
    store: true
  }
});

type WorkflowInput = { input_as_text: string };

// Main code entrypoint
export const runWorkflow = async (workflow: WorkflowInput) => {
  const state = {

  };
  const conversationHistory: AgentInputItem[] = [
    {
      role: "user",
      content: [
        {
          type: "input_text",
          text: workflow.input_as_text
        }
      ]
    }
  ];
  const runner = new Runner({
    traceMetadata: {
      __trace_source__: "agent-builder",
      workflow_id: "<固有のワークフローID>"
    }
  });
  const webResearchAgentResultTemp = await runner.run(
    webResearchAgent,
    [
      ...conversationHistory
    ]
  );
  conversationHistory.push(...webResearchAgentResultTemp.newItems.map((item) => item.rawItem));

  if (!webResearchAgentResultTemp.finalOutput) {
      throw new Error("Agent result is undefined");
  }

  const webResearchAgentResult = {
    output_text: JSON.stringify(webResearchAgentResultTemp.finalOutput),
    output_parsed: webResearchAgentResultTemp.finalOutput
  };
  const summarizeAndDisplayResultTemp = await runner.run(
    summarizeAndDisplay,
    [
      ...conversationHistory
    ]
  );

  if (!summarizeAndDisplayResultTemp.finalOutput) {
      throw new Error("Agent result is undefined");
  }

  const summarizeAndDisplayResult = {
    output_text: JSON.stringify(summarizeAndDisplayResultTemp.finalOutput),
    output_parsed: summarizeAndDisplayResultTemp.finalOutput
  };
}

  const summarizeAndDisplayResult = {
    output_text: JSON.stringify(summarizeAndDisplayResultTemp.finalOutput),
    output_parsed: summarizeAndDisplayResultTemp.finalOutput
  };
  console.log("webResearchAgentResult", webResearchAgentResult);
}

// Anthropicについて聞くよう設定、コマンドラインから受け取れば自由に修正可能
runWorkflow({ input_as_text: "Anthropic" });

{
  "name": "openai-agent-kit-test",
  "version": "1.0.0",
  "description": "OpenAI Agent Kit test project",
  "type": "module",
  "scripts": {
    "build": "tsc",
    "start": "node agent-builder-test.js"
  },
  "dependencies": {
    "@openai/agents": "^0.1.0",
    "zod": "^3.23.8",
    "dotenv": "^17.2.3"
  },
  "devDependencies": {
    "@types/node": "^22.0.0",
    "typescript": "^5.5.0"
  }
}

{
  "compilerOptions": {
    "target": "ES2022",
    "module": "ESNext",
    "moduleResolution": "bundler",
    "esModuleInterop": true,
    "strict": true,
    "skipLibCheck": true,
    "resolveJsonModule": true,
    "outDir": "./",
    "rootDir": "./"
  },
  "include": ["*.ts"],
  "exclude": ["node_modules"]
}

% npm run build

> openai-agent-kit-test@1.0.0 build
> tsc

% openai-agent-kit-test % npm start    

> openai-agent-kit-test@1.0.0 start
> node agent-builder-test.js

webResearchAgentResult {
  output_text: '{"companies":[{"company_name":"Anthropic","industry":"Artificial intelligence / AI research (safety-focused LLMs)","headquarters_location":"San Francisco, California, United States","company_size":"501–1,000 employees (approximate; rapidly growing)","website":"https://www.anthropic.com/","description":"Anthropic is an AI research and safety company that develops large language models and related tools with a focus on robustness, interpretability, and alignment. Known for its Claude family of conversational AI models, Anthropic emphasizes safety-first approaches, red-teaming, and policy/research work intended to reduce risks from advanced AI systems. The company partners with cloud providers and enterprise customers to offer API access and managed deployments of its models.","founded_year":2021}]}',
  output_parsed: { companies: [ [Object] ] }
}

cd backend
uv sync
export OPENAI_API_KEY=sk-proj-...
uv run uvicorn app.main:app --reload --port 8000

cd frontend
npm install
npm run dev