Project Positioning
Hello-Agents is a systematic agent learning tutorial launched by the Datawhale community, titled "Building Agents from Scratch," positioned as a from-scratch tutorial on agent principles and practice. The project divides Agent building into two schools: one is software engineering Agents such as Dify, Coze, and n8n, which are essentially process-driven software development with the LLM as a data processing backend; the other is AI-native Agents, that is, Agents truly driven by AI. This tutorial aims to lead learners to deeply understand and build the latter, namely true AI Native Agents.
Main Capabilities and Learning Content
The tutorial claims to help learners see through the surface of frameworks, start from the core principles of agents, go deep into the core architecture, understand classic paradigms, and ultimately build their own multi-agent applications. Specific capability points include:
- Understand core principles: deeply understand the concepts, history, and classic paradigms of agents.
- Hands-on implementation: master the use of popular low-code platforms and agent code frameworks.
- Self-developed framework: build the agent framework HelloAgents from scratch based on the OpenAI native API.
- Master advanced skills: gradually implement systematic technologies such as context engineering, Memory, protocols, and evaluation.
- Model training: master Agentic RL, with full-process practical training of LLMs from SFT to GRPO.
- Drive real cases: practically develop comprehensive projects such as an intelligent travel assistant and a cyber town.
- Job interviews: study interview questions related to agent job seeking.
Usage and Content Structure
The tutorial provides two online reading methods, overseas access and domestic acceleration, allowing learning without downloading. If local reading or contributing content is needed, the project provides a learning guide. The PDF version is completely open source and free, and has been pre-added with a Datawhale open source logo watermark that does not affect reading, to prevent it from being watermarked and sold.
The project is divided into five major parts:
- Part One: Foundations of Agents and Language Models (Chapters 1 to 3), starting from the definition, types, and development history of agents, and consolidating core knowledge of large language models.
- Part Two: Build Your Large Language Model Agent (Chapters 4 to 7), hands-on implementing classic paradigms such as ReAct, experiencing low-code platforms such as Coze, mastering mainstream frameworks such as LangGraph, and ultimately building your own agent framework from scratch.
- Part Three: Advanced Knowledge Expansion (Chapters 8 to 12), using the self-developed framework to explore memory and retrieval, context engineering,Core technologies such as agent training, learn communication protocols between multiple agents, and master methods for evaluating agent system performance.
- Part Four: Advanced Comprehensive Cases (Chapters 13 to 15), build an intelligent travel assistant, an automated deep research agent, and a cyber town that simulates social dynamics.
- Part Five: Graduation Project and Future Outlook (Chapter 16), build a complete multi-agent application, and look ahead to the future directions of agents.
The project's code folder provides all accompanying code. The tutorial recommends combining theory with practice, personally running, debugging, and even modifying the code. If you encounter problems, you can ask in the issue section. The community blog section collects learning insights and practice summaries, and submissions can be made to Extra-Chapter.
Applicable Scenarios and Target Audience
The project is suitable for AI developers, software engineers, students, and self-learners with a strong interest in cutting-edge AI technology who have a certain programming foundation. Before learning, you need basic Python programming ability and a basic conceptual understanding of large language models, for example, knowing how to call an LLM through an API. The project focuses on application and construction, and does not require a deep background in algorithms or model training.
Resources and Updates
- GitHub repository:https://github.com/datawhalechina/hello-agents
- PDF tutorial:https://github.com/datawhalechina/hello-agents/releases/latest/
- PDF domestic download address:https://www.datawhale.cn/learn/summary/239
Video courses are being released gradually. The HelloAgents framework has been updated to version V1.0.0 and will continue to be improved, adding more useful and lightweight tools and features, compatible with the learning version. In May 2026, the latest learning roadmap for agents was updated; click to check. The project has reached 70,000 stars and provides a survey questionnaire to collect learning needs. The follow-up work is "Training Agents from Scratch".
Contributions and License
The project is an open-source community, and contributions such as reporting bugs, making suggestions, improving content, and sharing practices are welcome. The tutorial uses the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. The project lists the responsible persons, co-initiators, content contributors, operational organizers, advising experts, and Extra-Chapter contributors. For detailed video links and accompanying materials, see the Hello-Agents video course recording co-creation.
Limitations and Not Yet Provided
Information such as the project's fee standards, prices, commercialization services, specific platform compatibility, and performance promises is not yet provided in the materials. Apart from the above PDF download address and domestic learning page, information on other platforms or channels is not yet provided.
