Bilibili AI Agent Development Tutorial: A Practical Guide to LangChain and MCP

Tutorial Overview and Background

Recently, Bilibili launched what is described as "the most comprehensive and detailed" zero-foundation AI Agent development tutorial, taught by Baoge (Mr. Wu) from Mashi Education. This tutorial is the first day of a four-day training camp, focusing on hands-on, zero-foundation practice with LangChain and MCP (Model Context Protocol). According to the instructor, this is part of their VIP course, designed to help learners quickly get started with AI Agent development.

The tutorial emphasizes "learn and get hired," and promises free materials including resume templates, interview guides, and project documentation. From the live interaction, audience familiarity with large models and LangChain varied, and the instructor adjusted the pace based on feedback.

LangChain and MCP: Core Concepts and Practical Value

LangChain is an open-source framework for building large model applications, while MCP is a standardized protocol designed to make it easier for large models to connect to external tools and data sources. The tutorial combines the two, demonstrating how to integrate MCP within LangChain to extend the boundaries of Agent capabilities.

For zero-foundation learners, understanding the role of MCP is particularly key: it acts like a "universal interface" that allows Agents to call external services such as search, databases, and APIs. During the live stream, the instructor stated that this content is one of the core tool courses they carefully prepared, covering an important technology stack for current large model application development.

LangChain's core abstractions include Chain (composable sequences of calls), Agent (autonomously selects tools based on goals), Tool (external function wrappers), Memory (conversation state management), and Retriever (retrieval augmentation). Among these, the Agent relies on the reasoning ability of the large model to decide which tool to call and in what order, while LangChain is responsible for describing tools as schemas the model can understand and parsing its output. Understanding this layering is crucial: the framework itself does not generate intelligence; it provides the engineering skeleton that glues model capabilities to external systems.

MCP (Model Context Protocol) was proposed by Anthropic at the end of 2024 and adopts a client-server architecture: the MCP Server exposes Resources, Tools, and Prompts, and the MCP Client (usually the Agent runtime) communicates with the server via JSON-RPC. Compared to hardcoding tools in code, MCP allows tool capabilities to be reused by different models and applications, forming an ecosystem-like "tool marketplace." Integrating MCP in LangChain typically involves using an adapter to convert MCP tools into LangChain Tool objects, thereby accessing standardized external capabilities while preserving the original Agent logic.

Four-Day Training Camp Curriculum and Supporting Resources

The training camp lasts four days. On the first day, Baoge explains LangChain and MCP; on subsequent days, Mr. Chen covers large model training-related content. Each day's live stream includes a lottery session with prizes such as cash red envelopes and shopping cards, aimed at increasing participant engagement.

After completing the four-day course, participants can receive a large model job-seeking resume template, a "Golden September and Silver October" interview guide, and Harness architecture project documentation. These resources are provided free of charge, but require continuous attendance and joining the study group. The instructor emphasizes that these materials are directly helpful for job hunting and project practice.

"Golden September and Silver October" refers to the autumn recruitment peak in China's job market every September and October, a window when companies release positions intensively and fresh graduates and job switchers submit applications densely. Therefore, interview question bank materials are in highest demand during this period. Resume templates and interview guides are essentially job-seeking auxiliary materials, and their value depends on whether they match the actual technology stack of the target position—for AI Agent development roles, employers typically value runnable projects, understanding of tool-calling pipelines, and debugging and troubleshooting experience more than memorizing question banks.

As for the "Harness architecture project documentation," judging by the name, it should refer to engineering architecture documentation based on a Harness-type toolchain or some internal naming. There is no unified standard definition in public materials. After receiving it, learners should first confirm its technology stack, dependency versions, and reproducibility before deciding whether to invest time in studying it.

Target Audience and Learning Advice

The tutorial is positioned as zero-foundation, but the live stream showed that about half of the audience already had some foundation. For complete beginners, it is recommended to first understand basic concepts of large models; for experienced learners, they can focus on MCP integration and advanced LangChain usage.

During learning, it is recommended to reproduce the code hands-on and use the documentation and project cases provided by the instructor to deepen understanding. In addition, although the training camp's interaction and lottery mechanisms can add fun, the core still lies in digesting and practicing the technical content.

For zero-foundation learners, when reproducing hands-on, it is recommended to follow the path of "get it running first, then break it down, then modify it": first, use a minimal example to run through a complete Agent call chain (model → select tool → execute → feed back results), confirming that the environment, API Key, and dependency versions are correct; second, break down the prompt template, tool descriptions, and parsing logic section by section to understand why the model made that decision; third, replace or add a custom tool and observe changes in Agent behavior. During debugging, enable LangChain's verbose or LangSmith tracing to view intermediate steps in full; otherwise, when tool calls fail, it is difficult to locate whether the problem lies in the schema description, the model output format, or the external interface.

Industry Observation: The Current State of AI Agent Development Training

With the explosion of large model applications, AI Agent development has become a hot skill, and various training courses are emerging endlessly. This tutorial uses "the most comprehensive and detailed" as its selling point and combines practical training with employment orientation, reflecting the market's urgent demand for application-oriented talent. However, learners need to carefully discern course quality and prioritize content with complete project cases and continuous updates.

From a technology trend perspective, the combination of LangChain and MCP is becoming one of the mainstream approaches for building complex Agents. Mastering these tools will help developers quickly implement AI applications in real business scenarios.