PaperOrchestra open source: multi-agent collaboration to write academic papers, the quality of literature review exceeds that of a single agent by 68%

Core conclusion

Writing academic papers is never as simple as "letting AI generate it in one breath". The Google research team proposed the PaperOrchestra multi-Agent framework to prove that an Agent collaboration system with clear division of labor can improve the quality of literature reviews by 50-68% and the overall paper quality by 14-38%. **On May 17, this framework was officially open sourced as a Skill package and is compatible with any programming agent such as Claude Code, Cursor, and Cline. **

Key Points

• Release form: Claude Code/Codex Skill package (pure instructions + deterministic scripts, zero API Key dependency)
• Core architecture: 5 Agent + 1 Orchestrator, Outline → Chart → Documentation → Writing → Refinement
• Literature Review Agent: Web Search + Semantic Scholar Verification + 90%+ Citation Coverage
• Paper quality assessment: 50-68% literature review winning rate, 14-38% overall quality winning rate
• GitHub star number: 488⭐ will be broken upon release, and the Chinese community will follow up simultaneously

Background: Dilemma of academic paper automation

AI writing papers is nothing new, but existing solutions have three fatal flaws:

**First, the quality of the literature review is dismal. ** Most AI writing tools only "generate text that looks like a literature review", with extremely low citation coverage of actual research progress. The PaperWritingBench benchmark found that more than 60% of citations in a literature review generated by a single-agent baseline were hallucinations.

**Second, the framework coupling is too tight. ** The existing automated writing system is deeply bound to specific experimental pipelines - if the research direction is changed, the entire solution will be useless.

**Third, the cost of in-depth participation is high. ** Fully automated generation solutions are either too simple (poor quality) or require complex API integration and model deployment.

PaperOrchestra's paper (arXiv:2604.05018) aims to overcome these three problems.

Detailed explanation of PaperOrchestra architecture

This system divides paper writing into 5 professional agents, each responsible for a link:

Key design: Step 2 and Step 3 are executed in parallel and do not block each other. The overall pipeline goes from original research materials to submittable LaTeX papers with zero manual intervention.

Literature Review Agent’s Secret Weapon

The biggest problem with traditional AI writing tools is "fabricated citations" - the Semantic Scholar API call results do not verify similarity, and the hallucination rate is high. PaperOrchestra’s Literature Review Agent does three things:

1. Web Search Candidate — Use the host Agent’s built-in search tool to discover relevant papers
2. Semantic Scholar Verification — Levenshtein The citation must be confirmed only if the similarity is > 70%
3. Automatic deduplication — Ensure that the same source is not cited repeatedly

Result: 90%+ citation coverage. This is a quantum leap in the academic writing scene.

Why this is good news for content creators too

On the surface, PaperOrchestra is an academic tool, but the design philosophy behind it deserves the attention of everyone who is doing AI automated content production:

• Professional Agent division of labor is better than large models working alone — This is true for writing papers, and it is also true for writing long articles for public accounts. One Agent is in charge of research, one is in charge of structure, and another is in charge of polishing. The effect is far better than one Prompt doing it all.
• Quality gate control is the core — PaperOrchestra’s refinement agent is essentially an automated quality gate. This is completely consistent with the AI ​​automated workflow plus quality gate concept we introduced before.
• Zero API Key Design — Pure Skill instructions + host Agent’s own capabilities, no third-party dependencies. This lowers the barriers to building AI automated systems.

Some content creators have already practiced a similar Agent division of labor model: a team used n8n + OpenAI to build an AI content automation website with a monthly income of $4,500. The core is to split "research → writing → publishing" into independent Agent division of labor and collaboration.

How to get started

Installation of PaperOrchestra's Skill package is extremely simple:

# Claude Code
claude mcp add paper-orchestra --path /path/to/PaperOrchestra

# GitHub
git clone https://github.com/Ar9av/PaperOrchestra.git
cd PaperOrchestra

# Claude Code
claude
> /skill load paper-orchestra
> /run paper-orchestra

• API Key
• Python （）
• Claude Code / Cursor

PaperOrchestra Claude Code、Cursor、n8n、OpenAI、Gemini. These tool entries appear naturally in the text, and the platform side will match the maintained tool library to generate floating cards.

Reference resources

• PaperOrchestra GitHub
• arXiv
• PaperWritingBench

Next action

If you are using AI for content production, PaperOrchestra’s multi-agent division of labor concept can be reused in your workflow:

• Want to build your own multi-agent automation system? Watch this tutorial: n8n + OpenAI build automated AI content distribution
• A team has already verified that this method of playing can make money: Use n8n + OpenAI to build an AI content website with monthly income $4,500
• Want to learn Claude Code from scratch? Take a look first: Claude Code content automation tutorial