Paper Research Agent

Paper Research Agent - Autonomous Multi-Agent Research System

When to Use

Use this skill when the user wants to:

Core Workflow

The system autonomously executes the full research pipeline:

User Query → Research Probe → PDF Download → Parallel Agent Analysis → Integrated Report

Phase 1: Research Probe (Automated)

Phase 2: PDF Download (Automated)

Phase 3: Parallel Agent Analysis (Automated - Key)

Phase 4: Report Integration (Automated)

Agent Analysis Requirements

Each sub-agent MUST generate a 6-section report following the detailed standards in: references/analysis_standards.md

SubAgent MUST read this reference file before starting analysis to understand:

Summary of 6 Required Sections

Section 1: Research Background

Section 2: Research Problem

Section 3: Core Innovation

Section 4: Experimental Design

Section 5: Key Insights

Section 6: Future Work

For full details, sub-section hints, and quality standards - READ references/analysis_standards.md

Quality Enforcement

Agents MUST:

Usage

Agent Execution (When User Requests Research)

Trigger phrases:

Agent Action:

Step 1: Execute main pipeline

import subprocess
result = subprocess.run([
    "python3", 
    "~/.openclaw/workspace/skills/paper-research-agent/scripts/research_pipeline.py",
    "--query", "{user_topic}",
    "--mode", "vertical",
    "--max-papers", "10",
    "--output", "./research_{topic}"
], capture_output=True, text=True)
print(result.stdout)

Step 2: Read generated agent tasks

import json
with open("./research_{topic}/_agent_tasks.json") as f:
    tasks = json.load(f)

Step 3: Spawn parallel sub-agents for analysis (CRITICAL)

# Spawn multiple agents in parallel for each paper
for task_info in tasks:
    sessions_spawn(
        agentId="main",
        mode="run", 
        runtime="subagent",
        task=task_info['task'],
        timeoutSeconds=600  # 10 minutes per paper
    )

Important: Launch as many agents in parallel as possible for speed.

Step 4: After all agents complete, integrate results

# Collect all analysis reports
Generate integrated survey
Present to user

Output Structure

research_output/
├── _research_summary.json              # Research metadata
├── probe/
│   ├── _probe_results.json            # Search results
│   └── _probe_report.md               # Human-readable probe report
├── papers/
│   ├── {title}-{arxiv_id}.pdf         # Downloaded PDFs
│   └── ...
├── analysis/
│   ├── {title}-{arxiv_id}_analysis.md # 6-section agent reports
│   └── ...
└── _integrated_survey.md              # Final integrated survey

Key Scripts

Report Format Standards

Each sub-agent analysis report MUST follow this exact 6-section structure:

# 📄 {Paper Title}
> ArXiv ID: {id}  
> Authors: {authors}  
> Published: {date}
Section 1: Research Background
Domain context
Key prior works (3-5 papers with citations)
Technical state at publication time
Citations: [Section X.Y]
Section 2: Research Problem
SPECIFIC problem being solved
SPECIFIC limitations of existing methods (quote original)
Core assumptions
Citations: [Section X.Y, "exact quote"]
Section 3: Core Innovation
Method/system architecture (detailed)
Technical details (network structure, dimensions)
Key formulas in LaTeX: $...$
Comparison table:
  | Aspect | Prior Work | This Paper | Advantage |
  |--------|-----------|------------|-----------|
What is genuinely new
Section 4: Experimental Design
Dataset: Name, size, characteristics
Baseline methods: Specific names
Metrics: Formulas, units
Results table (REAL data):
  | Method | Metric1 | Metric2 |
  |--------|---------|---------|
  | This | X.XX | X.XX |
  | Baseline | X.XX | X.XX |
Ablation study results
Section 5: Key Insights
Core findings from experiments
What works/doesn't work
Design choices and impact
Practical recommendations
Section 6: Future Work
Limitations acknowledged by authors
Unsolved problems
Future directions (3+)
*Analysis by Paper Research Agent*  
*Date: {timestamp}*

Quality Requirements:

Error Handling

If paper download fails:

If agent analysis fails:

Best Practices

1. For deep research: Use --mode vertical (searches 4 levels deep) 2. For exploration: Use --mode iterative (progressive discovery) 3. For specific paper: Use --mode horizontal (find related work) 4. Parallel agents: System auto-spawns optimal number based on paper count 5. Quality check: Always verify a few random citations manually

Example Session

User: "帮我调研扩散策略在机器人操作中的应用"

Agent: 1. Executes research probe with query "扩散策略 机器人操作" 2. Finds 30 related papers across 4 levels 3. Downloads PDFs for top 10 papers 4. Spawns 10 sub-agents in parallel 5. Each agent analyzes one paper with 6-section format 6. Collects all analyses 7. Generates integrated survey with comparison tables 8. Presents final report to user

Output: Complete research package with all papers analyzed and integrated survey.

Dependencies

Required Python packages (auto-installed):

Notes

Agent Execution (When User Requests Research)

Trigger phrases:

Agent Action:

Step 1: Execute main pipeline

import subprocess
result = subprocess.run([
    "python3", 
    "~/.openclaw/workspace/skills/paper-research-agent/scripts/research_pipeline.py",
    "--query", "{user_topic}",
    "--mode", "vertical",
    "--max-papers", "10",
    "--output", "./research_{topic}"
], capture_output=True, text=True)
print(result.stdout)

Step 2: Read generated agent tasks

import json
with open("./research_{topic}/_agent_tasks.json") as f:
    tasks = json.load(f)

Step 3: Spawn parallel sub-agents for analysis (CRITICAL)

# Spawn multiple agents in parallel for each paper
for task_info in tasks:
    sessions_spawn(
        agentId="main",
        mode="run", 
        runtime="subagent",
        task=task_info['task'],
        timeoutSeconds=600  # 10 minutes per paper
    )

Important: Launch as many agents in parallel as possible for speed.

Step 4: After all agents complete, integrate results

# Collect all analysis reports
Generate integrated survey
Present to user

1. For deep research: Use --mode vertical (searches 4 levels deep) 2. For exploration: Use --mode iterative (progressive discovery) 3. For specific paper: Use --mode horizontal (find related work) 4. Parallel agents: System auto-spawns optimal number based on paper count 5. Quality check: Always verify a few random citations manually