Hello-Agents: A Systematic, Hands-On Guide to Building Agents from Scratch

Background and Context

If 2024 marked the year of the great model wars, 2025 is being described as the year of the agent, with technical focus shifting from training ever-larger foundation models toward building smarter agent applications. Yet systematic, practice-oriented tutorials remain extremely scarce, and this gap forms the direct motivation behind Hello-Agents, an open-source tutorial produced by the Datawhale community, which has accumulated tens of thousands of stars. The project deliberately positions itself between theory and practice: it refuses to merely teach learners how to click through a low-code interface, and it also declines to dwell on abstract concepts from papers. Instead, it aims to be a hands-on guide that transforms users of large language models into builders of agent systems.

Within the agent ecosystem, the tutorial distinguishes two schools of construction. One camp comprises software-engineering-style agents such as Dify, Coze, and n8n, which are fundamentally process-driven development where the LLM serves as a data-processing backend. The other camp consists of AI-native agents, driven genuinely by AI. Hello-Agents explicitly commits to the latter, urging readers to look past the surface of frameworks and return to the core principles and classic paradigms of agents, ultimately building their own multi-agent application by hand.

Deep Analysis

The tutorial organizes its content into four sections of increasing depth. The first covers agent definitions, types, and paradigms, tracing the evolution from symbolism to LLM-driven agents, while grounding readers in Transformers, prompt engineering, and the limitations of mainstream LLMs. The second moves into hands-on work, implementing ReAct, Plan-and-Solve, and Reflection, while comparing low-code platforms such as Coze, Dify, and n8n against frameworks like AutoGen, AgentScope, and LangGraph to help readers build selection judgment.

A standout is Chapter 7, which builds a custom agent framework called HelloAgents from scratch using the native OpenAI API, letting readers see exactly how a framework operates internally. The third section covers systematic advanced skills: memory systems with RAG, context engineering for continuous interaction, communication protocols including MCP, A2A, and ANP, and a full pipeline of Agentic-RL training from SFT to GRPO, concluding with core metrics, benchmarks, and evaluation frameworks. The fourth section ties everything together through综合 projects such as an AI travel assistant and a cyber town.

The material is organized as Markdown documents in a docs directory, spanning from a preface through twelve chapters with clear navigation, and most sections are tagged with completion status so readers can jump or read systematically. The tutorial offers both online and local reading paths, with separate addresses for international access and domestic acceleration, lowering the entry barrier. However, because it covers deep topics like Agentic-RL and GRPO, readers still need a working knowledge of Python and large language models to fully absorb the advanced material.

Industry Impact

The value of Hello-Agents extends beyond knowledge transfer to shaping an engineering mindset that understands essence rather than depending on black boxes. It helps developers avoid becoming mere callers of low-code platforms or framework APIs, instead mastering the general principles of agent design, a capability that is particularly scarce amid rapid technical iteration. For engineering teams, systematically studying context engineering, memory, protocols, and evaluation supports the construction of more reliable, evaluable multi-agent applications.

The tutorial's differentiated stance—insisting on understanding core mechanisms rather than merely operating framework interfaces—sets it apart from the many tutorials that simply搬运 material. Its complete loop from principle to framework, from component to system, and from training to evaluation is precisely what distinguishes it from scattered blog posts. The emphasis on hands-on practice, backed by runnable code examples and a self-developed framework project, reinforces this builder-oriented philosophy.

Outlook

Potential risks remain present. The agent domain's frameworks and protocols evolve extremely fast, so specific framework usage in the tutorial may become outdated as versions advance, requiring readers to cross-check against official documentation. Meanwhile, Agentic-RL training demands significant compute and engineering capability, meaning ordinary learners may not be able to fully reproduce those segments.

Directions worth watching include whether the tutorial can keep pace with the ecological evolution of protocols such as A2A and MCP, and whether the Agentic-RL content can further lower its practical barrier so more developers can run the complete training chain from SFT to GRPO. Overall, Hello-Agents has become a strong entry point into the agent world for developers, researchers, and learners seeking to evolve from users into builders.