Dispatching Parallel Agents
Coordinate multiple AI agents to execute tasks simultaneously, maximizing development velocity while preventing conflicts. This is the core skill for multi-agent collaboration.
When to Use
- •2+ independent tasks can be worked on simultaneously
- •No shared state between tasks (different files/modules)
- •Total work > 2 hours and parallelization is beneficial
- •Overnight execution where multiple agents can run unattended
Agent Capability Matrix
Model Backgrounds & Strengths
| Agent | Model | Background | Best For | Cost | Speed |
|---|---|---|---|---|---|
| Antigravity | Claude Opus 4.5 | Deep reasoning, planning | Architecture, coordination, complex debugging | 💰💰💰 | 🐢 |
| Amp | Claude Sonnet 4.5 | Accuracy, low hallucination | Code review, quality assurance | 💰💰 | 🐇 |
| Codex | GPT-5.2 Codex | Patient execution | Detailed refactoring, overnight tasks, React | 💰💰 | 🐢 |
| GLM | GLM-4.7 | Speed, Chinese, cost | Large code generation, repetitive tasks | 💰 | 🐇🐇 |
| MiniMax | MiniMax | Type-safe, structured | Theme systems, type definitions | 💰 | 🐇 |
| Grok | xAI Grok | Real-time search | Research, API docs, best practices | 💰💰 | 🐇 |
Performance Benchmarks (2025-2026)
code
GLM-4.7: - SWE-bench: 73.8% - LiveCodeBench-v6: 84.9% (open-source #1) - Cost: ~15% of Claude Codex (GPT-5.2): - First-try success: 37% - With retries: 70.2% - Development speedup: 2-3x Claude Sonnet 4.5: - Accuracy: Highest tier - Hallucination rate: Lowest - Large codebase understanding: Excellent
RALPH Loop Integration
The RALPH loop (Read → Analyze → Learn → Plan → Hypothesize) integrates with multi-agent workflow:
code
┌─────────────────────────────────────────────────────────────────────┐ │ RALPH LOOP │ ├─────────────────────────────────────────────────────────────────────┤ │ │ │ ┌──────┐ ┌─────────┐ ┌───────┐ ┌──────┐ ┌────────────┐ │ │ │ READ │ → │ ANALYZE │ → │ LEARN │ → │ PLAN │ → │ HYPOTHESIZE│ │ │ └──────┘ └─────────┘ └───────┘ └──────┘ └────────────┘ │ │ ↑ │ │ │ └─────────────── Feedback Loop ──────────────────────┘ │ │ │ ├─────────────────────────────────────────────────────────────────────┤ │ MULTI-AGENT DISPATCH PHASE │ ├─────────────────────────────────────────────────────────────────────┤ │ │ │ After PLAN: Antigravity decomposes into parallel tasks │ │ │ │ ┌────────────────┐ ┌────────────────┐ ┌────────────────┐ │ │ │ Agent A (Codex)│ │ Agent B (GLM) │ │ Agent C (Mini) │ │ │ │ Task Guide A │ │ Task Guide B │ │ Task Guide C │ │ │ │ Files: A/* │ │ Files: B/* │ │ Files: C/* │ │ │ └───────┬────────┘ └───────┬────────┘ └───────┬────────┘ │ │ │ │ │ │ │ └────────────────────┼────────────────────┘ │ │ ↓ │ │ ┌──────────────────┐ │ │ │ Antigravity │ │ │ │ Reviews & Merges │ │ │ └──────────────────┘ │ │ ↓ │ │ Back to READ (next iteration) │ │ │ └─────────────────────────────────────────────────────────────────────┘
Conflict Prevention Strategies
1. File-Level Isolation
Principle: Each agent owns exclusive files/directories.
code
❌ BAD: Both agents edit same file Agent A: components/Header.tsx (lines 1-50) Agent B: components/Header.tsx (lines 51-100) Result: Merge conflicts guaranteed ✅ GOOD: Agents own different files Agent A: components/Header.tsx, api/stocks.ts Agent B: themes/*.ts, stores/theme.store.ts Result: Clean merge
2. Dependency Ordering
Principle: Infrastructure before features, types before implementation.
code
Phase 1 (Foundation) - Must complete first: └── types/theme.ts [GLM] └── api/client.ts [Codex] Phase 2 (Implementation) - Can parallelize after Phase 1: ├── themes/dark.theme.ts [GLM] ├── themes/pixel.theme.ts [GLM] ├── api/stocks.ts [Codex] └── api/watchlist.ts [Codex] Phase 3 (Integration) - After Phase 2: └── components/ThemeSwitcher.tsx [Any]
3. Contract-First Development
Principle: Define interfaces before implementation.
typescript
// Define contract FIRST (in shared types file)
interface Theme {
id: string;
name: string;
colors: ThemeColors;
}
// Then agents implement independently
// Agent A: implements getTheme(id) → Theme
// Agent B: implements <ThemeSwitcher themes={Theme[]} />
4. Task Boundary Markers
Include clear ownership in task guides:
markdown
### Task: Theme Store Implementation **Owner**: MiniMax **Files OWNED** (exclusive write): - stores/theme.store.ts - themes/registry.ts **Files READ-ONLY** (for reference): - types/theme.ts - tokens.css **Do NOT touch**: - components/* (Codex owns these)
Real-World Example: DragonFly Parallel Execution
Scenario: Codex + MiniMax Working Simultaneously
code
┌──────────────────────────────────────────────────────────────────┐ │ TIME: 2026-01-20 15:17 │ │ User: "让 Codex 干 UI 优化，让 MiniMax 干主题系统" │ ├──────────────────────────────────────────────────────────────────┤ │ │ │ ┌─────────────────────────────────┐ │ │ │ CODEX: UI-POLISH-COMBINED.md │ │ │ │ ─────────────────────────────── │ │ │ │ Owns: │ │ │ │ - components/LeftPane.tsx │ │ │ │ - components/StockChart.tsx │ │ │ │ - components/RightPane.tsx │ │ │ │ - styles/tokens.css (shadows) │ │ │ │ │ │ │ │ Tasks: │ │ │ │ ✅ BF-001: Name display fix │ │ │ │ ✅ UI-001: Sidebar modernize │ │ │ │ ✅ UI-003: Chart controls │ │ │ │ ✅ UI-005: AI input glow │ │ │ └─────────────────────────────────┘ │ │ ↓ │ │ (Running ~2.5h) │ │ ↓ │ │ ┌─────────────────────────────────┐ │ │ │ MINIMAX: GLM-P1-THEME-GUIDE.md │ │ │ │ ─────────────────────────────── │ │ │ │ Owns: │ │ │ │ - types/theme.ts │ │ │ │ - themes/*.theme.ts │ │ │ │ - themes/registry.ts │ │ │ │ - stores/theme.store.ts │ │ │ │ - hooks/useTheme.ts │ │ │ │ │ │ │ │ Tasks: │ │ │ │ ✅ G-001: Theme types │ │ │ │ ✅ G-002-005: 4 themes │ │ │ │ ✅ G-008: Zustand store │ │ │ │ ✅ G-010: ThemeSwitcher │ │ │ └─────────────────────────────────┘ │ │ ↓ │ │ (Running ~2.5h) │ │ ↓ │ ├──────────────────────────────────────────────────────────────────┤ │ TIME: 2026-01-20 19:17 (4 hours later) │ │ │ │ RESULT: Both complete, pnpm check passes ✅ │ │ No merge conflicts │ │ 4 themes working + UI improvements │ │ │ │ NEXT: Antigravity reviews → Creates Round 2 fixes │ └──────────────────────────────────────────────────────────────────┘
Post-Merge Review Findings
After parallel execution, Antigravity reviews and creates new tasks:
markdown
## Review Findings → Round 2 Tasks | Issue Found | New Task | Assignee | |-------------|----------|----------| | Search dropdown behavior | FIX-001: Keyboard navigation | Codex | | Button position blocks price | FIX-002: Move to left toolbar | Codex | | Period buttons are mock | FIX-003: Connect to backend | Codex | | Cyberpunk theme glitch | FIX-005: Border rendering | MiniMax | | Dark theme input colors | FIX-006: Use CSS variables | Codex |
Task Guide Template for Parallel Execution
markdown
# [Task Group Name] > **Executor**: Codex / GLM / MiniMax > **Estimated Time**: X hours > **Parallel Group**: A / B / C ← Indicates parallel execution group --- ## ⚠️ Ownership Declaration ### Files I OWN (exclusive write access): - `path/to/file1.ts` - `path/to/file2.tsx` - `path/to/directory/*` ### Files I READ (no modifications): - `types/*.ts` - `styles/tokens.css` ### Files FORBIDDEN (another agent owns): - `themes/*` (MiniMax owns) - `api/*` (Codex owns) --- ## Tasks ### T-001: First Task ... --- ## Verification \`\`\`bash pnpm check # Type safety pnpm dev # Visual check \`\`\` ## Problem Log If blocked, document here immediately: - [x] Problem: ... - [x] Workaround: ... - [ ] Needs human decision: ...
Dispatch Decision Tree
code
┌─────────────────┐
│ New Task Set │
└────────┬────────┘
│
┌────────▼────────┐
│ Can parallelize? │
└────────┬────────┘
│
┌────────────────┼────────────────┐
│ │ │
YES DEPENDS NO
│ │ │
┌─────────▼─────────┐ │ ┌────────▼────────┐
│ Check file overlap │ │ │ Sequential exec │
└─────────┬─────────┘ │ │ (use best agent)│
│ │ └─────────────────┘
┌───────┴───────┐ │
│ │ │
NONE OVERLAP │
│ │ │
▼ ▼ ▼
┌───────────┐ ┌───────────┐ ┌───────────┐
│ Dispatch │ │ Split by │ │ Phase into│
│ to agents │ │ files/dirs│ │ sequential│
└───────────┘ └───────────┘ └───────────┘
Best Practices
- •Always declare file ownership in task guides
- •Types/interfaces first to establish contracts
- •No shared mutable state between parallel agents
- •Batch reviews after parallel work completes
- •Quick iteration cycles - don't let issues accumulate
- •Visual verification with agent-browser for UI work
- •pnpm check gate - mandatory before considering done
- •Problem logging - agents must document blockers immediately
Anti-Patterns to Avoid
code
❌ Two agents editing same component ❌ Agent A depends on Agent B's in-progress work ❌ No ownership declaration in task guides ❌ Skipping type definitions before implementation ❌ Long parallel runs without checkpoints ❌ No verification commands in task guides
Communication Templates
Dispatching to Agent
code
请阅读并执行 tasks/epics/{epic}/codex/TASK-GUIDE.md
这是与 {other-agent} 并行执行的任务。
你负责的文件: {list}
禁止修改: {other-agent's files}
Agent Status Report
markdown
## Status: [Agent Name] ### Completed - ✅ T-001: Description - ✅ T-002: Description ### In Progress - 🔄 T-003: Description (eta: 30min) ### Blocked - ⚠️ T-004: Need backend API for X ### Files Modified - `path/to/file1.ts` (new) - `path/to/file2.tsx` (modified) ### Verification \`\`\` pnpm check → ✅ Pass \`\`\`