使用監督 Agent 建立研究協調器

在本指南中，你會使用監督 Agent 建立研究協調器，以協調多個專門 Agent。協調器會將研究工作委派給研究 Agent，並將寫作工作委派給寫作 Agent，然後把結果整合成詳細報告。

你會設定角色清晰的子 Agent，並設定監督 Agent 來協調它們。你亦會使用委派 hook 控制執行，並使用評分器驗證工作是否完成。

先決條件

• 已安裝 Node.js v22.13.0 或更新版本
• 受支援 Model Provider 的 API 金鑰
• 現有的 Mastra 項目（按照安裝指南設定新項目）

建立研究 Agent

研究 Agent 專門收集任何主題的事實資料。它會以精簡項目符號摘要傳回重要事實及來源。

建立新檔案 src/mastra/agents/research-agent.ts：

src/mastra/agents/research-agent.ts

import { Agent } from '@mastra/core/agent'

export const researchAgent = new Agent({
  id: 'research-agent',
  name: 'Research Specialist',
  description:
    'Specializes in gathering factual information and data on any topic. ' +
    'Returns concise bullet-point summaries with key facts and sources. ' +
    'Does not write full articles or narrative content.',
  instructions:
    'You are a research specialist. When given a topic, gather key facts, ' +
    'statistics, and information. Present findings as clear bullet points. ' +
    'Include sources when possible. Focus on accuracy and completeness.',
  model: 'openai/gpt-5-mini',
})

description 欄位非常重要，它有助監督 Agent 判斷何時應委派給此 Agent。清晰的描述可提高委派準確度。

建立寫作 Agent

寫作 Agent 會將研究資料轉化為結構良好、段落完整且行文流暢的文章。

建立新檔案 src/mastra/agents/writing-agent.ts：

src/mastra/agents/writing-agent.ts

import { Agent } from '@mastra/core/agent'

export const writingAgent = new Agent({
  id: 'writing-agent',
  name: 'Writing Specialist',
  description:
    'Transforms research material into well-structured written content. ' +
    'Produces full paragraphs and complete articles with proper flow. ' +
    'Best used after research has been gathered.',
  instructions:
    'You are a writing specialist. Transform research and information into ' +
    'well-written articles. Use complete paragraphs, clear structure, and ' +
    'engaging language. Maintain a professional yet accessible tone. ' +
    'Ensure the content flows naturally from introduction to conclusion.',
  model: 'openai/gpt-5-mini',
})

建立監督 Agent

監督 Agent 會協調研究及寫作工作。其指示會定義委派策略，包括何時使用各個子 Agent，以及如何整合結果。Memory 直接在 Agent 上設定。

建立新檔案 src/mastra/agents/supervisor-agent.ts：

src/mastra/agents/supervisor-agent.ts

import { Agent } from '@mastra/core/agent'
import { Memory } from '@mastra/memory'
import { LibSQLStore } from '@mastra/libsql'
import { researchAgent } from './research-agent'
import { writingAgent } from './writing-agent'

export const supervisorAgent = new Agent({
  id: 'supervisor-agent',
  name: 'Research Coordinator',
  instructions: `You coordinate research and writing tasks using specialized agents.

Available resources:
- research-agent: Gathers factual data and sources (returns bullet points)
- writing-agent: Transforms research into well-structured articles (returns full paragraphs)

Delegation strategy:
1. For research requests: Delegate to research-agent first to gather facts
2. For writing requests: Delegate to writing-agent with any available research context
3. For comprehensive reports: Delegate to research-agent first, then writing-agent
4. Always ensure you have gathered sufficient information before producing final output

Success criteria:
- All aspects of the user's request are addressed
- Information is accurate and well-sourced
- Final output is well-formatted and complete
- If anything is missing or uncertain, continue gathering information`,
  model: 'openai/gpt-5.6-sol',
  agents: {
    researchAgent,
    writingAgent,
  },
  memory: new Memory({
    storage: new LibSQLStore({
      id: 'mastra-storage',
      url: 'file:mastra.db',
    }),
  }),
  defaultOptions: {
    maxSteps: 10,

    // Monitor progress after each iteration
    onIterationComplete: async context => {
      console.log(`\n✓ Iteration ${context.iteration} complete`)
      console.log(`  Finish reason: ${context.finishReason}`)
      console.log(`  Response length: ${context.text.length} chars\n`)

      // Continue until task is complete
      return { continue: true }
    },

    // Control delegations
    delegation: {
      onDelegationStart: async context => {
        console.log(`→ Delegating to: ${context.primitiveId}`)

        // Add context for specific agents
        if (context.primitiveId === 'research-agent') {
          return {
            proceed: true,
            modifiedPrompt: `${context.prompt}\n\nFocus on recent developments (2024-2025) and include statistics.`,
          }
        }

        return { proceed: true }
      },

      onDelegationComplete: async context => {
        console.log(`✓ Completed: ${context.primitiveId}\n`)

        // Handle errors: bail to stop execution and provide feedback
        if (context.error) {
          console.error('Delegation failed:', context.error)
          context.bail() // Stop further delegations
          return {
            feedback: `Delegation to ${context.primitiveId} failed: ${context.error}. Try a different approach.`,
          }
        }
      },

      // Only pass last 10 messages to subagents
      messageFilter: ({ messages }) => {
        return messages.slice(-10)
      },
    },
  },
})

在 src/mastra/index.ts 中向 Mastra 註冊監督 Agent：

src/mastra/index.ts

import { Mastra } from '@mastra/core'
import { supervisorAgent } from './agents/supervisor-agent'

export const mastra = new Mastra({
  agents: { supervisorAgent },
})

監督 Agent 上的 defaultOptions 會設定委派 hook 及反覆運算監察：

• onDelegationStart 修改研究 Agent 的提示，以要求近期資料
• onDelegationComplete 記錄完成狀態，在發生錯誤時停止，然後提供回饋
• messageFilter 將上下文限制為最後 10 則訊息，以提高效率
• onIterationComplete 在每次反覆運算後監察進度

測試基本監督 Agent

在 src/index.ts 建立檔案，以便與監督 Agent 互動：

src/index.ts

import { supervisorAgent } from './mastra/agents/supervisor-agent'

async function main() {
  const topic = 'artificial intelligence in education'
  console.log(`\nTopic: ${topic}\n`)

  const stream = await supervisorAgent.stream(
    `Research ${topic} and write a comprehensive article about it`,
  )

  // Stream the response
  console.log('📝 Final Report:\n')
  for await (const chunk of stream.textStream) {
    process.stdout.write(chunk)
  }
  console.log('\n')
}

main()

監督 Agent 會使用 defaultOptions 中設定的委派 hook 及反覆運算監察。

加入工作完成評分

工作完成評分器會自動驗證工作是否完成，避免監督 Agent 過早結束。

在 src/mastra/scorers/task-complete-scorer.ts 中建立評分器：

src/mastra/scorers/task-complete-scorer.ts

import { createScorer } from '@mastra/core/evals'

export const taskCompleteScorer = createScorer({
  id: 'task-complete',
  name: 'Task Completeness',
  description: 'Checks if the research and writing task has been fully completed',
}).generateScore(async context => {
  const text = (context.run.output || '').toString()

  // Check if response contains required elements
  const hasSubstantialContent = text.length > 500
  const hasStructure = text.includes('\n\n') // Multiple paragraphs
  const hasContext = /\d{4}/.test(text) // Contains years/dates

  // Return 1 if complete, 0 if not
  if (hasSubstantialContent && hasStructure && hasContext) {
    return 1
  }

  return 0
})

安裝 Evals 依賴套件：

npm

npm install @mastra/evals

pnpm

pnpm add @mastra/evals

Yarn

yarn add @mastra/evals

Bun

bun add @mastra/evals

更新 src/mastra/agents/supervisor-agent.ts 中的監督 Agent，在 defaultOptions 內加入工作完成評分：

src/mastra/agents/supervisor-agent.ts

import { Agent } from '@mastra/core/agent'
import { Memory } from '@mastra/memory'
import { LibSQLStore } from '@mastra/libsql'
import { researchAgent } from './research-agent'
import { writingAgent } from './writing-agent'
import { taskCompleteScorer } from '../scorers/task-complete-scorer'

export const supervisorAgent = new Agent({
  id: 'supervisor-agent',
  name: 'Research Coordinator',
  instructions: `You coordinate research and writing tasks using specialized agents.

Available resources:
- research-agent: Gathers factual data and sources (returns bullet points)
- writing-agent: Transforms research into well-structured articles (returns full paragraphs)

Delegation strategy:
1. For research requests: Delegate to research-agent first to gather facts
2. For writing requests: Delegate to writing-agent with any available research context
3. For comprehensive reports: Delegate to research-agent first, then writing-agent
4. Always ensure you have gathered sufficient information before producing final output

Success criteria:
- All aspects of the user's request are addressed
- Information is accurate and well-sourced
- Final output is well-formatted and complete
- If anything is missing or uncertain, continue gathering information`,
  model: 'openai/gpt-5.6-sol',
  agents: {
    researchAgent,
    writingAgent,
  },
  memory: new Memory({
    storage: new LibSQLStore({
      id: 'mastra-storage',
      url: 'file:mastra.db',
    }),
  }),
  defaultOptions: {
    maxSteps: 10,

    onIterationComplete: async context => {
      console.log(`\n✓ Iteration ${context.iteration} complete`)
      console.log(`  Finish reason: ${context.finishReason}`)
      console.log(`  Response length: ${context.text.length} chars\n`)
      return { continue: true }
    },

    delegation: {
      onDelegationStart: async context => {
        console.log(`→ Delegating to: ${context.primitiveId}`)

        if (context.primitiveId === 'research-agent') {
          return {
            proceed: true,
            modifiedPrompt: `${context.prompt}\n\nFocus on recent developments (2024-2025) and include statistics.`,
          }
        }

        return { proceed: true }
      },

      onDelegationComplete: async context => {
        console.log(`✓ Completed: ${context.primitiveId}\n`)

        if (context.error) {
          console.error('Delegation failed:', context.error)
          context.bail() // Stop further delegations
          return {
            feedback: `Delegation to ${context.primitiveId} failed: ${context.error}. Try a different approach.`,
          }
        }
      },

      messageFilter: ({ messages }) => {
        return messages.slice(-10)
      },
    },

    // Validate task completion
    isTaskComplete: {
      scorers: [taskCompleteScorer],
      strategy: 'all',
      onComplete: async result => {
        console.log('\n🎯 Completion Check:')
        console.log(`  Complete: ${result.complete}`)
        console.log(`  Score: ${result.scorers[0]?.score}\n`)
      },
    },
  },
})

評分器會檢查內容是否充實、結構是否恰當，以及是否包含上下文資料。如果工作尚未完成，監督 Agent 會繼續反覆運算。現在，所有 hook 及工作完成評分都已在 Agent 的 defaultOptions 中設定，因此會自動套用至每次呼叫。

測試研究協調器

執行協調器以查看實際運作：

npx tsx src/index.ts

你會看到監督 Agent 先委派給研究 Agent，再委派給寫作 Agent，而記錄會顯示委派流程：

Topic: artificial intelligence in education

→ Delegating to: research-agent
✓ Iteration 1 complete
  Finish reason: tool-calls
  Response length: 0 chars

✓ Completed: research-agent

→ Delegating to: writing-agent
✓ Iteration 2 complete
  Finish reason: tool-calls
  Response length: 0 chars

✓ Completed: writing-agent

🎯 Completion Check:
  Complete: true
  Score: 1

✓ Iteration 3 complete
  Finish reason: stop
  Response length: 1247 chars

📝 Final Report:

Artificial Intelligence in Education: Transforming Learning in 2024-2025

[The coordinator will produce a comprehensive article combining research findings with well-structured writing...]

由於 Agent 回應並非確定，你的輸出可能有所不同，但委派模式會保持一致。

後續步驟

你可以擴充此研究協調器，以：

• 加入更多專門 Agent（事實查核員、編輯、引用格式化工具）
• 為品質指標（可讀性、來源品質）實作自訂評分器
• 加入用於網絡搜尋或資料庫存取的 Tool
• 為複雜的多步驟研究程序建立 Workflow
• 使用結構化輸出產生特定格式的報告

了解更多：

• 監督 Agent
• Agent.stream() 參考