Building a Multi-Agent AI System from Scratch: MCP Practice of Deep Research Agent and Writing Workflow

This article introduces a hands-on workshop project that guides developers to build a multi-agent AI system from scratch, which includes a deep research agent and a writing workflow—both served as MCP (Model Context Protocol) servers. Through complete code and materials, the project helps developers master key challenges such as architectural design and component collaboration, and has important reference value for AI agent development.

With the evolution of large language models, AI agents are moving toward practical applications, but building reliable multi-agent systems faces challenges such as architecture and collaboration. The project is named "Designing Real-World AI Agents Workshop" and is positioned as a hands-on resource, providing code, presentations, and videos with the goal of enabling developers to build runnable systems. The core components are the deep research agent (responsible for information collection and analysis) and the writing workflow (generating structured content), both served as MCP servers, reflecting the modularization trend.

MCP is an open protocol proposed by Anthropic that standardizes the interaction between AI models and external tools. The advantages of encapsulating components as MCP servers include: interoperability (supports calls from clients like Claude Desktop), separation of concerns (transparent internal details), flexibility in composition (components can be combined like building blocks), and security control (permission and resource management).

The deep research agent is the core of information processing and needs to address: 1. Information retrieval strategies (search engine integration, document parsing, knowledge base querying); 2. Information synthesis and verification (multi-source cross-validation, citation tracking, iterative refinement); 3. Structured output (outlines, key point lists, etc., for easy downstream consumption).

The writing workflow transforms research results into content through phased processing: planning (determine structure and arguments), drafting (generate first draft), revision (review and polish), and finalization (format adjustment). It supports human-machine collaboration, introducing human review at key decision points, and realizes interaction through MCP tool calls.

The project demonstrates two collaboration architectures:

1. Pipeline collaboration (the output of the research agent directly serves as the input of the writing workflow—clear structure but limited flexibility);
2. Feedback loop (the writing workflow calls back the research agent to supplement information, which is closer to the human team mode). The project code supports the basic pipeline mode and reserves space for extending complex interactions.

The practical value of the project lies in allowing participants to learn experiences such as architectural trade-offs, error handling, performance optimization, and test evaluation. Application scenarios include content marketing automation, academic research assistance, enterprise knowledge management, and rapid news media reporting. The MCP-based architecture can be combined with other tools to build complex systems.

This project provides high-quality practical references for AI agent developers, demonstrating a modular and composable way to build multi-agent systems. It is of great value to technical personnel who want to delve into AI agent architecture, MCP applications, or content automation. The community can contribute by expanding information sources, enhancing style control, adding evaluation frameworks, developing visual interfaces, etc.