research-orchestrator

Research Orchestrator

Transform OpenClaw into a Manus-level deep research agent with parallel sub-agents, iterative search, and professional report generation.

Features

• 🤖 Parallel Sub-Agents: Use sessions_spawn for concurrent research
• 🔄 Iterative Search: Search → Analyze gaps → Search again
• 📊 Multi-Source: Web search + academic sources
• ✅ Fact Verification: Cross-validation with confidence
• 📄 Professional Reports: Markdown + PDF with styling
• 🌍 Multi-Language: Auto-detect user language

Trigger Conditions

• "Deep research on..." / "深度研究..."
• "Comprehensive analysis of..." / "全面分析..."
• "Research report about..." / "...研究报告"
• "Market analysis" / "市场分析"
• "帮我研究一下..."

---

Execution Workflow

When user requests deep research, follow this exact workflow:

Step 1: Create Task Structure

TASK_ID=$(date +%Y%m%d_%H%M%S)
WORKSPACE="${OPENCLAW_WORKSPACE:-$PWD}"
TASK_DIR="$WORKSPACE/research-orchestrator/tasks/$TASK_ID"

mkdir -p "$TASK_DIR/research"
mkdir -p "$TASK_DIR/analysis"
mkdir -p "$TASK_DIR/output"

echo "📋 Task created: $TASK_ID"
echo "📁 Workspace: $TASK_DIR"

Step 2: Analyze & Decompose Task

Based on user query, identify 3-5 research angles. Example for "AI芯片市场2026":

Research Angles:
1. Market Size & Growth - 市场规模与增长
2. Key Players - 主要厂商分析
3. Technology Trends - 技术发展趋势
4. Investment & M&A - 投资与并购
5. Policy Environment - 政策环境

Save plan to $TASK_DIR/plan.md.

Step 3: Launch Parallel Sub-Agents

CRITICAL: Use sessions_spawn to create parallel research agents.

For each research angle, spawn a sub-agent:

sessions_spawn with:
- session: "research-$TASK_ID-{angle_id}"
- mode: "run"
- task: Research instructions for this angle
- streamTo: "parent" (for progress visibility)

Research Agent Instructions Template:

You are a research agent. Your task: {angle_name}

Search Queries:
{queries}

Instructions:
1. Use web-search skill to search each query
2. For each result, extract key facts
3. Note source URL and assess credibility
4. Save findings to: $TASK_DIR/research/{angle_id}.md

Format findings as:
## {angle_name}

### Key Findings
- Finding 1 [Source: url] (Confidence: High)
- Finding 2 [Source: url] (Confidence: Medium)

### Data Points
| Metric | Value | Source | Confidence |
|--------|-------|--------|------------|

### Sources
1. [Source Name](url) - Credibility: High

Step 4: Iterative Search (Gap Analysis)

After initial research completes, analyze gaps:

# Analyze what's missing
gaps = identify_gaps(task_dir)

# For each gap, spawn another search
for gap in gaps:
    sessions_spawn with additional queries

Step 5: Fact Verification

Cross-verify key claims:

For each major finding:
1. Check if multiple sources agree
2. Assess source credibility
3. Flag conflicting information
4. Assign confidence score

Step 6: Deep Analysis

Synthesize research into insights:

Analysis should include:
1. Key Insights - Most important findings
2. Trend Analysis - Where things are heading
3. Risk Assessment - What could go wrong
4. Opportunities - Where to focus

Step 7: Generate Report

Create professional report with:

# Generate Markdown report
python3 $TASK_DIR/../scripts/report_generator.py generate $TASK_DIR {lang}

# Convert to PDF
python3 $TASK_DIR/../scripts/pdf_converter.py convert $TASK_DIR/output/report.md

---

Real Execution Example

User: "帮我深度研究一下2026年AI芯片市场"

Agent should:

1. Create task structure
   mkdir -p tasks/20260323_100000/{research,analysis,output}

2. Decompose into angles
   - market_size, competitors, technology, investment

3. Launch 4 parallel sub-agents using sessions_spawn
   
   Agent 1 (market_size):
   "Search: 'AI chip market size 2026', '半导体市场规模预测'
    Find market size data, growth rates, forecasts
    Save to tasks/.../research/market_size.md"
   
   Agent 2 (competitors):
   "Search: 'NVIDIA AMD Intel AI chips market share'
    Analyze key players, their strategies
    Save to tasks/.../research/competitors.md"
   
   Agent 3 (technology):
   "Search: 'AI chip architecture trends 2026', 'GPU NPU发展'
    Identify technology trends
    Save to tasks/.../research/technology.md"
   
   Agent 4 (investment):
   "Search: 'AI chip investment 2026', 'AI芯片投融资'
    Track investment and M&A activity
    Save to tasks/.../research/investment.md"

4. Wait for agents to complete (stream progress)

5. Verify facts across sources

6. Generate analysis report

7. Create final report (MD + PDF)

8. Output:
   📄 report.md
   📄 report.pdf

---

Progress Tracking

Show progress during execution:

📊 Research Progress
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Task: AI芯片市场2026
Status: Researching

Phase: Parallel Research
├─ market_size: ✅ Done (8 sources)
├─ competitors: 🔄 Running (5 sources)
├─ technology: ⏳ Pending
└─ investment: ⏳ Pending

Total Sources: 13
Elapsed: 3m 45s

---

Multi-Language Support

Output language matches user input:

• User writes in Chinese → Output Chinese report
• User writes in English → Output English report
• User specifies "用英文输出" → Output in specified language

---

Key Differences from Basic Skills

Basic Skill	Research Orchestrator
Single execution	Parallel sub-agents
One search	Iterative multi-search
Simple output	Professional reports
No verification	Fact cross-checking
Static results	Dynamic gap analysis

---

Dependencies

• python3 - Script execution
• curl - API calls
• node - PDF generation (md-to-pdf)
• web-search skill - Web searching

---

Notes

• Research typically takes 5-15 minutes
• Parallel agents provide 3-4x speedup
• Reports include source citations
• Professional PDF formatting included
• Supports 50+ languages