Antigravity マルチエージェント開発 実践設計パターン — AgentKit で構築する自律型AIチーム

# AGENTS.md — Antigravity エージェント定義ファイル
version: "2.0"
 
agents:
  # === Code Generation & Refactoring ===
  code-writer:
    description: "新規コード生成・機能実装"
    model: "gemini-3.1-pro"
    temperature: 0.7
    max_tokens: 4000
 
  code-reviewer:
    description: "コード品質監査・リファクタリング提案"
    model: "gemini-3.1-pro"
    temperature: 0.3
    max_tokens: 2000
 
  refactoring-expert:
    description: "大規模リファクタリング・設計改善"
    model: "gemini-3.1-pro"
    temperature: 0.5
    max_tokens: 3000
 
  # === Testing & Quality ===
  test-generator:
    description: "ユニット・統合テスト自動生成"
    model: "gemini-3.1-flash"
    temperature: 0.3
    max_tokens: 2000
 
  test-reviewer:
    description: "テストケース品質監査"
    model: "gemini-3.1-pro"
    temperature: 0.4
    max_tokens: 1500
 
  # === Architecture & Design ===
  architect:
    description: "システムアーキテクチャ設計"
    model: "gemini-3.1-pro"
    temperature: 0.5
    max_tokens: 5000
 
  performance-analyzer:
    description: "パフォーマンスプロファイリング・最適化"
    model: "gemini-3.1-pro"
    temperature: 0.4
    max_tokens: 3000
 
  # === Security & Compliance ===
  security-auditor:
    description: "セキュリティ脆弱性診断"
    model: "gemini-3.1-pro"
    temperature: 0.2
    max_tokens: 2500
 
  compliance-checker:
    description: "規制・ライセンス準拠チェック"
    model: "gemini-3.1-flash"
    temperature: 0.3
    max_tokens: 1500
 
  # === Documentation ===
  documentation-writer:
    description: "API・技術ドキュメント自動生成"
    model: "gemini-3.1-flash"
    temperature: 0.4
    max_tokens: 3000
 
  # === Data & Deployment ===
  data-analyzer:
    description: "データ変換・スキーマ最適化"
    model: "gemini-3.1-pro"
    temperature: 0.5
    max_tokens: 2500
 
  devops-engineer:
    description: "CI/CD・デプロイメント最適化"
    model: "gemini-3.1-flash"
    temperature: 0.4
    max_tokens: 2000
 
  # === Communication ===
  technical-writer:
    description: "ユーザー向けドキュメント・ガイド作成"
    model: "gemini-3.1-flash"
    temperature: 0.6
    max_tokens: 3000
 
  code-explainer:
    description: "複雑なコードの説明・学習ガイド"
    model: "gemini-3.1-flash"
    temperature: 0.6
    max_tokens: 2000
 
  # === Specialist (Additional) ===
  langsmith-optimizer:
    description: "LLMチェーン・プロンプト最適化"
    model: "gemini-3.1-pro"
    temperature: 0.3
    max_tokens: 1500
 
  cost-analyzer:
    description: "トークン・コスト分析・予算管理"
    model: "gemini-3.1-flash"
    temperature: 0.2
    max_tokens: 1000

# AGENTS.md — Sequential Pattern
orchestration:
  pattern: "sequential"
  name: "Code Quality Pipeline"
 
  steps:
    - agent: "code-writer"
      input: "Feature request: User authentication"
      output_var: "generated_code"
 
    - agent: "code-reviewer"
      input: "{{ generated_code }}"
      output_var: "reviewed_code"
      condition: "quality_score > 7"
 
    - agent: "test-generator"
      input: "{{ reviewed_code }}"
      output_var: "test_suite"
 
    - agent: "security-auditor"
      input: "{{ reviewed_code }}"
      output_var: "security_report"
 
  # ===== CRITICAL: 暴走防止設定 =====
  max_iterations: 3           # 最大3ステップまで（無限ループ防止）
  loop_detection: true        # ループ自動検出
  human_approval: true        # 各ステップで人間確認
  timeout_seconds: 300        # タイムアウト5分
  max_tokens_per_step: 4000   # ステップあたりの最大トークン
  total_token_budget: 20000   # パイプライン全体の予算

antigravity orchestrate --config AGENTS.md --pattern sequential
# 出力:
# Step 1: code-writer → generated_code.ts (2104 tokens)
# [PENDING APPROVAL] Review generated code? (y/n)
# > y
# Step 2: code-reviewer → reviewed_code.ts (1852 tokens)
# ...

orchestration:
  pattern: "parallel"
  name: "Multi-Perspective Code Audit"
 
  parallel_tasks:
    - agent: "code-reviewer"
      input: "{{ source_code }}"
      output_var: "quality_report"
 
    - agent: "security-auditor"
      input: "{{ source_code }}"
      output_var: "security_report"
 
    - agent: "performance-analyzer"
      input: "{{ source_code }}"
      output_var: "perf_report"
 
  # 全タスク完了を待つ
  wait_for_all: true
 
  # === Cost Control ===
  parallel_token_limit: 3000  # タスクあたり最大3000トークン
  max_concurrent_tasks: 3     # 同時実行数3

const results = await orchestrator.parallel([
  { agent: 'code-reviewer', input: code },
  { agent: 'security-auditor', input: code },
  { agent: 'performance-analyzer', input: code },
]);
 
// 結果統合
const synthesis = await orchestrator.synthesize(results);
// → { quality: 8.2, security: 9.1, performance: 7.5 }

orchestration:
  pattern: "router"
  name: "Smart Task Router"
 
  router_logic:
    - condition: "code_language == 'typescript'"
      agent: "code-writer"
      config:
        model: "gemini-3.1-pro"
 
    - condition: "code_language == 'python'"
      agent: "code-writer"
      config:
        model: "gemini-3.1-pro"
        temperature: 0.6
 
    - condition: "task_type == 'security_audit'"
      agent: "security-auditor"
 
    - condition: "task_type == 'performance'"
      agent: "performance-analyzer"
 
    # Fallback
    - default: "code-writer"
 
  # ===== Safety Rails =====
  max_routing_depth: 2        # 最大2段階のルーティング
  fallback_timeout: 60        # 判断が遅れたら60秒でフォールバック

orchestration:
  pattern: "orchestrator-worker"
  name: "Project Manager Pattern"
 
  orchestrator:
    agent: "architect"  # マネージャー役
    instruction: |
      You are the Project Manager. Decompose the task:
      1. Identify subtasks
      2. Assign to appropriate worker agents
      3. Aggregate results
      4. Ensure quality standards
 
  workers:
    - "code-writer"
    - "test-generator"
    - "documentation-writer"
    - "security-auditor"
 
  # === Task Breakdown ===
  max_subtasks: 5              # 最大5つのサブタスク
  worker_timeout_per_task: 180 # 1タスク3分以内
 
  # === Feedback Loop ===
  quality_threshold: 8.0       # 品質スコア8以上で完了
  max_iterations: 2            # 最大2回の改善ループ

orchestration:
  pattern: "swarm"
  name: "Collaborative AI Swarm"
 
  swarm_config:
    swarm_size: 4            # 4エージェント
    swarm_members:
      - "code-writer"
      - "code-reviewer"
      - "architect"
      - "test-generator"
 
    # === Communication Protocol ===
    consensus_required: true # 意思決定は合意ベース
    voting_strategy: "majority"  # 過半数で決定
    max_consensus_rounds: 3
 
    # === Evolutionary Improvement ===
    iteration_count: 5       # 5ラウンド反復
    fitness_function: "code_quality_and_test_coverage"
 
    # === Token Management ===
    swarm_token_budget: 50000  # スウォーム全体のバジェット
    token_per_agent: 10000     # エージェントあたり最大10K
 
    # === Loop Prevention ===
    detect_stagnation: true  # 進捗がなければ停止
    stagnation_threshold: 3  # 3ラウンド改善なしで終了

// Manager Surface Configuration
const managerSurface = {
  title: "Antigravity Development Team",
  layout: "grid",
 
  agents_panel: [
    {
      name: "code-writer",
      status: "idle",
      current_task: "Waiting for assignment",
      tokens_used: 0,
      efficiency: 0.0,
    },
    {
      name: "code-reviewer",
      status: "working",
      current_task: "Auditing UserAuthService.ts",
      tokens_used: 1852,
      efficiency: 92.3,
    },
    // ... (other agents)
  ],
 
  metrics_panel: {
    total_tokens_budget: 20000,
    tokens_used: 5234,
    tokens_remaining: 14766,
 
    team_progress: 65,  // %
    quality_score: 8.4,
    cost_efficiency: 0.78,  // (quality / tokens_used)
  },
 
  timeline_view: {
    stages: [
      { name: "Code Generation", duration: "12s", status: "✅ complete" },
      { name: "Code Review", duration: "8s", status: "⏳ in-progress" },
      { name: "Testing", duration: "pending", status: "⏱️ queued" },
    ]
  },
 
  controls: {
    pause_all: false,
    emergency_stop: false,
    adjust_budget: true,
    override_agent: ["code-reviewer"],  // どのエージェントでも人間が介入可能
  }
};

safety_rails:
  # === Loop Detection ===
  loop_detection:
    enabled: true
 
    # 検出パターン
    detect_patterns:
      - agent_output_matches_input:
          threshold: 0.95  # 95%以上類似 → ループ判定
          description: "Agent produced same output"
 
      - same_agent_called_twice:
          max_occurrences: 1  # 同一エージェントは最大1回まで
          description: "Agent can't improve further"
 
      - no_progress_after_iterations:
          iterations: 3
          quality_delta: 0.1  # 品質向上が0.1以下 → 停止
          description: "Quality improvement stalled"
 
    # 検出時の処理
    on_loop_detected:
      action: "pause"  # "pause" / "fallback" / "stop"
      human_review: true
      message: |
        ⚠️ LOOP DETECTED: architect agent has run 4 times
        without meaningful improvement.
 
        Options:
        1. [R]esume with different approach
        2. [S]witch to different agent
        3. [A]ccept current result
        4. [C]ancel operation

// Iteration limit + token budget combo
const safeguards = {
  max_iterations: 5,           // 最大5ループ
  max_tokens_per_iteration: 4000,  // 1回あたり4K tokens
  total_token_budget: 30000,   // 全体で30K tokens
 
  // === Hard Stops ===
  timeout_minutes: 10,         // 10分でタイムアウト
 
  // === Soft Stops (人間判断） ===
  warn_at_80_percent_budget: true,  // 80%で警告
  require_approval_after_iteration_3: true,  // 3回目以降は承認必須
 
  monitor: async (state) => {
    if (state.iterations >= 3) {
      // 第3イテレーション以降は人間承認が必須
      const approval = await requestHumanApproval({
        current_quality: state.quality_score,
        prev_quality: state.prev_quality_score,
        tokens_used: state.tokens_used,
        estimated_tokens_remaining: state.budget_remaining,
      });
 
      if (!approval) {
        return { action: 'HALT', reason: 'No human approval' };
      }
    }
 
    if (state.tokens_used > state.budget * 0.9) {
      // 予算の90%に達したら警告して停止確認
      console.warn(`⚠️ BUDGET WARNING: ${state.tokens_used} / ${state.budget}`);
    }
  }
};

# AGENTS.md — Production Template
version: "2.0"
name: "Enterprise SaaS Development Pipeline"
description: "Production-grade multi-agent orchestration for SaaS projects"
 
# === Agent Pool ===
agents:
  # Core development team
  code-writer:
    model: "gemini-3.1-pro"
    temperature: 0.7
    max_tokens: 4000
    expertise: "Full-stack implementation"
 
  code-reviewer:
    model: "gemini-3.1-pro"
    temperature: 0.3
    max_tokens: 2000
    expertise: "Code quality, design patterns"
 
  test-generator:
    model: "gemini-3.1-flash"  # Flash for speed
    temperature: 0.3
    max_tokens: 2000
    expertise: "Unit, integration, E2E tests"
 
  security-auditor:
    model: "gemini-3.1-pro"  # Pro for thoroughness
    temperature: 0.2
    max_tokens: 3000
    expertise: "OWASP, CVE scanning, compliance"
 
  architect:
    model: "gemini-3.1-pro"
    temperature: 0.5
    max_tokens: 5000
    expertise: "System design, scalability"
 
# === Orchestration ===
orchestration:
  pattern: "orchestrator-worker"
 
  orchestrator:
    agent: "architect"
    role: "Project Manager"
    instruction: |
      - Decompose feature requests into subtasks
      - Assign tasks to appropriate workers
      - Monitor quality and budget
      - Escalate blockers to human
 
  workflow:
    - name: "Feature Implementation"
      steps:
        - agent: "code-writer"
          input: "{{ feature_request }}"
          timeout: 120
 
        - agent: "code-reviewer"
          input: "{{ code_from_previous }}"
          quality_threshold: 7.5
 
        - agent: "test-generator"
          input: "{{ reviewed_code }}"
          coverage_threshold: 80
 
        - agent: "security-auditor"
          input: "{{ code_and_tests }}"
          risk_level: "low"  # Must be low or medium
 
# === Safety Rails ===
safety:
  max_iterations: 4
  loop_detection: true
  max_total_tokens: 50000
  max_per_step_tokens: 5000
  timeout_minutes: 30
  human_approval_required_after_iteration: 2
 
  # Cost alerts
  cost_alerts:
    warn_at_75_percent: true
    warn_at_90_percent: true
    hard_stop_at_100_percent: true
 
# === Monitoring & Logging ===
monitoring:
  log_all_agent_calls: true
  track_token_usage: true
  measure_quality_scores: true
 
  metrics:
    - code_quality_trend
    - test_coverage_trend
    - security_score
    - cost_per_feature
    - time_per_feature
    - agent_utilization
 
# === Manager Surface Config ===
ui:
  layout: "dashboard"
  refresh_interval: 2  # seconds
 
  panels:
    - agents_status
    - metrics_summary
    - task_timeline
    - token_usage_chart
    - quality_scorecard