PaperOrchestra: Reconstruction Practice of a Multi-Agent Automated Academic Paper Writing Framework

PaperOrchestra is a multi-agent framework targeting pain points in academic writing, aiming to automate the paper writing process through AI agent collaboration. The paperorchestra-for-codex introduced in this article is an independent reconstructed version of this framework, using GPT/Codex and the OMX toolchain to replace the original model stack while retaining core design concepts (explicit artifacts, verification gates, fidelity checks), providing researchers with a runnable academic writing assistant. The project is positioned as a 'paper drafting scaffold under human supervision' to help reduce transactional work while preserving space for human creativity.

The original PaperOrchestra paper proposed the framework concept of multi-agent collaborative paper writing, but the official implementation was not fully open-sourced. The kosh7707/paperorchestra-for-codex project chose to reconstruct independently, aiming to retain core design concepts while using a more open and accessible technology stack. Key challenges in reconstruction include: maintaining the semantic equivalence of prompts across different models, defining clear phase boundaries and artifact handover rules, and establishing reliable verification and review mechanisms.

The core design of PaperOrchestra revolves around three principles: 1. Explicit artifacts and phase boundaries: Each phase produces clear output artifacts as input for the next phase, with traceable and verifiable status; 2. Verification and review gates: After the phase ends, verification checks are performed to ensure the artifact quality meets standards before entering the next phase; 3. Fidelity checks: Built-in output quality checks, including verifiability of literature citations, correctness of charts, LaTeX compilation success rate, etc.

The system decomposes paper writing into six specialized phases: 1. Literature research and knowledge acquisition: Retrieve literature through Semantic Scholar, build a corpus, and support simulation mode; 2. Paper structure planning: Generate a chapter outline based on literature and user contribution descriptions; 3. Chapter-by-chapter content generation: Generate content chapter by chapter to ensure logical coherence and academic norms; 4. Chart and visualization generation: Support generating placeholders or graphics, with reserved extension points; 5. Citation management and literature arrangement: Automatically maintain BibTeX entries and format citations; 6. LaTeX compilation and artifact output: Compile into LaTeX documents and PDFs, and provide environment check tools.

The project's technical implementation includes: three layers of usage modes (safe local demonstration, real model call, OMX native mode) to adapt to different needs; deep integration of Codex CLI and OMX toolchain for easy command-line operation and workflow integration; artifact export mechanism supporting packaged output files; session state management that transparently displays progress and environment configuration, facilitating debugging.

The project is positioned as a 'paper drafting and reconstruction scaffold under human supervision', not a fully autonomous generator: 1. Draft generation rather than final draft delivery, providing first drafts, skeletons, and review suggestions; 2. Assisting research rather than replacing it, accelerating the expression process rather than thinking. Suitable scenarios: quickly generating initial draft frameworks, automatically organizing citations, generating standardized method/experiment structures, and academic writing teaching examples.

Current limitations: limited chart fidelity (mainly placeholders), technical claims requiring human review, citation quality not guaranteed to meet specific journal standards, non-official implementation. Future directions: enhance chart generation and integration with professional tools, expand support for multi-disciplinary writing norms, introduce human-machine collaboration interfaces, develop domain-specific templates, and deepen integration with academic search engines to verify citation quality.