Research Agent Orchestration System for Hermes

A multi-agent research workflow system where specialized sub-agents collaborate via a shared knowledge base to produce literature reviews and research proposals.

Architecture

hermes-research-agent/
├── README.md
├── requirements.txt
├── workflow.py              # Main workflow orchestrator
├── tools/
│   ├── __init__.py
│   ├── delegate_tools.py     # Shared KB tools for sub-agents
│   └── research_task_runner.py  # CLI entry point for sub-agents
├── roles/                   # Agent system prompts
│   ├── __init__.py
│   ├── literature_searcher.py
│   ├── paper_analyzer.py
│   ├── review_writer.py
│   └── idea_generator.py
├── shared/                   # Shared knowledge base (JSON files)
└── results/                 # Output results (one subdir per task)

How It Works

1. Main agent receives user request: "我想做关于 XXX 的文献综述"
2. Main agent spawns sub-agents using delegate_task
3. Sub-agents communicate via shared JSON files
4. Main agent coordinates execution order and synthesizes final output

Sub-Agent Roles

Agent	Role	Input	Output
Literature Search	Search academic databases	Topic	shared/literature_search.json
Paper Analyzer	Deep analysis of papers	shared/literature_search.json	shared/paper_analysis.json
Review Writer	Write structured review	shared/paper_analysis.json	results/{task_id}/literature_review.md
Idea Generator	Propose research directions	shared/paper_analysis.json	shared/proposed_ideas.json

Installation

pip install -r requirements.txt

Usage

Main Workflow (via Hermes)

User: "我想做关于 CRISPR 基因编辑的文献综述"
Main agent → spawns sub-agents in sequence

Testing Directly

# Run full workflow with a test topic
python workflow.py "lysozyme"

# Run individual sub-agents
python -m tools.research_task_runner literature_search "CRISPR gene editing"
python -m tools.research_task_runner paper_analyze
python -m tools.research_task_runner review_writer
python -m tools.research_task_runner idea_generator

Workflow Steps

1. Literature Search → Searches PubMed, Semantic Scholar, arXiv for relevant papers
2. Paper Analysis → Deep-dives into found papers, extracts themes and gaps
3. Review Writing → Synthesizes into structured Markdown review
4. Idea Generation → Proposes novel research directions based on gaps

Shared Knowledge Base

All sub-agents read/write JSON files in shared/:

• task_{task_id}_literature_search.json - Raw search results
• task_{task_id}_paper_analysis.json - Thematic analysis
• task_{task_id}_proposed_ideas.json - Research directions

Results are stored in results/{task_id}/.