Multi-Agent Writing Workflow: An OpenCode-Based Automatic Generation System for Chinese Reflective Essays

This article introduces the Reflective-Essay-Agent-Workflow project, an OpenCode-based multi-agent writing workflow that enables the automatic generation of high-quality Chinese reflective essays through coordinated collaboration among multiple specialized AI agents. The project demonstrates the innovative application of multi-agent systems in creative writing, providing a new technical paradigm for AI-assisted content creation.

With the improvement of large language model capabilities, AI-assisted writing has evolved from text completion to complex content generation. However, single models face issues such as context length limitations, difficulty maintaining style consistency, and insufficient logical coherence when handling long-form, multi-layered creative writing. Multi-agent systems address these by decomposing complex tasks into subtasks, with specialized agents collaborating to enhance professionalism in each link while maintaining overall consistency. This project is a practice of this concept in the field of Chinese reflective essays.

Reflective-Essay-Agent-Workflow is an open-source multi-agent writing framework designed for Chinese reflective essays (integrating personal insights, philosophical thinking, and literary expression). This genre requires deep theme understanding, elegant language expression, and coherent progression of ideas. The project is built on the OpenCode platform, using a file system as the communication mechanism between agents to achieve a decentralized and scalable writing workflow.

Core Concept

The core of the project is 'collaborative division of labor and iterative refinement', breaking down writing into links such as requirement understanding, outline planning, content generation, style polishing, logical verification, and final integration, each handled by a dedicated agent.

File-Based Communication Mechanism

Agents communicate via files in the tmp/ directory, with advantages including persistent state, decoupled collaboration, auditability, and human intervention. A typical workflow: user_input.txt → planner/ → outline.txt → writer/ → draft_v1.txt → editor/ → draft_v2.txt → reviewer/ → feedback.txt → ...

Agent Roles

• Planner: Converts user input into a structured outline, responsible for theme analysis, structure design, and material planning
• Writer: Generates text based on the outline, with capabilities in paragraph writing, opinion expression, and example selection
• Editor: Optimizes language, focusing on word choice, sentence structure adjustment, and rhetorical enhancement
• Reviewer: Evaluates article quality, checking logical consistency, theme focus, and structural integrity

OpenCode Platform Integration

Leverages OpenCode features such as Session management (independent agent sessions), file system API (standardized reading/writing), and task orchestration (multiple execution modes).

Prompt Engineering Strategies

System prompts follow principles of role immersion, input/output specifications, quality guidelines, and error handling. For example, the planner prompt clearly identifies it as a senior reflective essay planner, with output format as tmp/outline.json including title, introduction, body paragraphs, conclusion, etc.

Iterative Optimization Mechanism

• Inner Loop: Self-correction by individual agents (e.g., writers adjusting content that deviates from the outline)
• Outer Loop: Cross-agent collaborative optimization (e.g., reviewers triggering editors or writers to rework)

Typical Usage Flow

1. Theme Input: Users provide themes and requirements via user_input.txt
2. Automatic Execution: The system runs agents according to the preset workflow
3. Process Monitoring: Users view intermediate products in tmp/
4. Human Intervention: Pause at key nodes for modifications before continuing
5. Result Acquisition: The final essay is output to output/final_essay.txt

Applicable Themes

• Philosophical Thinking: Abstract topics such as the meaning of life and value choices
• Social Observation: Reflections and comments on contemporary social phenomena
• Cultural Criticism: Interpretations and insights on literary and artistic works
• Personal Narrative: Life insights combined with personal experiences

Limitations

• Factual accuracy requires manual verification
• Originality risks require plagiarism checks
• Multi-agent collaboration may lead to style inconsistencies
• Long articles (>5000 words) are difficult to control in structure

Advantages of Multi-Agent Systems

• Specialized Division of Labor: Enhances individual capabilities
• Interpretability: Transparent creation process facilitates problem localization
• Scalability: Easy to add new agents (e.g., fact-checkers)
• Human-Agent Collaboration: Humans can intervene at any stage

Industry Insights

This project represents the evolution of AI writing from 'single-agent operation' to 'team collaboration', raising thoughts on task decomposition schemes, communication protocol design, quality evaluation systems, and computing cost control.

Technical Aspects

1. Agent Learning: Continuous improvement from manual modifications
2. Dynamic Division of Labor: Automatically adjust agent configurations based on tasks
3. Parallel Optimization: Explore parallelizable writing links
4. Memory Mechanism: Introduce long-term memory to maintain style consistency

Application Aspects

1. Genre Expansion: Novels, poetry, scripts, etc.
2. Multi-Language Support: Adapt to English and other languages
3. Personalized Customization: Adjust style based on user preferences
4. Collaborative Writing: Mixed collaboration between multiple humans and agents