Blog-Writing-Agent: An Intelligent Blog Writing Agent Based on LangGraph

Blog-Writing-Agent is a multi-step intelligent blog writing agent built on LangGraph, designed to address the pain points of generative AI text such as lack of factual basis and superficial content. By introducing real-time web research, structured planning, and evidence-driven writing, it upgrades AI from a mere "text generator" to a "research-writing assistant", capable of producing high-quality blog content rooted in facts and with technical depth. Key features include adaptive writing mode routing, parallel chapter generation, structured evidence management, etc.

In the era of generative AI, text generated by LLMs often suffers from lack of factual basis and is filled with "AI-sounding" clichés, making it difficult to provide valuable technical insights. Blog-Writing-Agent was created to address this pain point: its core idea is to let AI "think" rather than just "write", generating in-depth, fact-supported content through multi-step workflows, real-time web research, and structured planning.

Blog-Writing-Agent uses LangGraph as its core orchestration framework (an extension of LangChain that supports complex control flows), consisting of 5 core nodes forming a directed graph:

1. Router: Determines whether research is needed and selects the mode (closed_book/hybrid/open_book);
2. Research: Collects evidence using the Tavily Search API;
3. Orchestrator: Generates a structured blog outline;
4. Worker: Executes chapter writing in parallel;
5. Reducer: Merges chapters into a complete article. This design leverages LangGraph's Send primitive to achieve parallelization and improve generation efficiency.

Intelligent Routing

According to the topic, it adaptively selects 3 modes:

• Closed Book: Suitable for basic concepts (e.g., Transformer principles), no research needed;
• Hybrid: Suitable for topics requiring updated examples, 45-day research window;
• Open Book: Suitable for time-sensitive topics (e.g., AI news), 7-day research window.

Research Layer

Uses Tavily Search to obtain structured results, converts raw results into EvidenceItem (including title, URL, date, etc.) via Pydantic, and performs deduplication based on URL and filtering by time window.

Orchestration and Writing

The Orchestrator generates a structured Plan (including chapter tasks, objectives, word count, etc.); Worker nodes write chapters in parallel, strictly following prompts (covering key points, controlling word count, citing evidence, etc.).

Tech Stack

• Core Framework: LangChain/LangGraph;
• LLM Interface: ChatOllama (default llama3.1);
• Search Tool: Tavily;
• Structured Validation: Pydantic;
• Language: Python3.10+.

State Management

Manages data flow via LangGraph's State (TypedDict), which includes fields like topic, mode, evidence, plan, sections, etc. The sections field uses Annotated + operator.add to enable automatic collection of parallel results.

Practical Insights

1. Fact First: Mandatory evidence citation and time-sensitive filtering solve the hallucination problem;
2. Modular Design: Each node has a single responsibility, facilitating testing and optimization;
3. Adaptive Strategy: Three modes cover different scenarios;
4. Parallel Optimization: Leverage LangGraph's fan-out capability to improve efficiency;
5. Structured Output: Pydantic ensures data consistency.

Application Scenarios

Suitable for generating content with high factual accuracy requirements, such as technical blogs, news summaries, research abstracts, etc. Extensible features include image generation, multilingual translation, SEO optimization, etc.

Blog-Writing-Agent demonstrates how to upgrade LLMs into "research-writing assistants". By using web research, structured planning, and evidence citation, it solves the problems of insufficient factuality and "AI-sounding" language in pure LLM writing. This dual-drive model of "thinking + writing" represents a promising direction for AI-assisted content creation.