Sixteen chapters form an agent curriculum
Hello-Agents is closer to an open textbook with labs than an installable product. Its 16 chapters begin with agent and language-model basics, then implement ReAct, Plan-and-Solve, and Reflection. Later chapters cover memory, retrieval, context engineering, MCP and other communication protocols, agent training, and evaluation. The final stretch builds a travel assistant, a deep-research agent, and a simulated town before asking the learner to complete a multi-agent capstone.
That breadth is the main reason to choose it. A reader can see low-code platforms, AutoGen, AgentScope, LangGraph, and a small homegrown framework in one sequence instead of learning one vendor's vocabulary first. The course assumes basic Python and prior knowledge of calling a language model through an API. It does not require a deep machine-learning background, although the Agentic RL chapter eventually reaches SFT and GRPO.
The English track covers the main 16 chapters
Chinese is the primary language, and the repository says so plainly. An English README links to translated versions of all 16 main chapters, including the projects and capstone. Release V1.0.3, published July 17, 2026, also includes Chinese and English PDFs. English readers therefore have a coherent route through the central book rather than a machine-translated landing page with dead chapter links.
The edges are less uniform. Community extras and co-creation projects retain many Chinese names and instructions, and the DataAnalysisAgent README we inspected is in Chinese. This matters because those additions are where readers encounter independent implementation choices and setup details. Someone who reads only English can finish the main course, but may need translation help when debugging a contributed project or following discussion in an issue.
What happened when we ran it
Our sandbox targeted Co-creation-projects/1zrj-DataAnalysisAgent/ at commit 45dd84e. Installation succeeded in 41 seconds, adding 137 packages and taking 316 MB on disk. The build completed in 4 seconds. Pip-audit reported 0 known vulnerabilities in the installed Python dependencies.
There was no tests script or target, so the test step was skipped. The 185.2 MB checkout contained 1,856 files and roughly 144,188 lines of source. Our scan found 0 CI workflow files, no Dockerfile, and no tests directory. Those facts make the boundary important: one nested project resolved and built, while neither its behavior nor the rest of the course passed a shared automated suite in our run.
One nested notebook needs an LLM API
The measured DataAnalysisAgent is a Jupyter project that reads tabular data, asks a language model for analysis and chart code, and writes an ECharts HTML file plus a Markdown report. Its README asks users to provide LLM_MODEL_ID, LLM_API_KEY, and LLM_BASE_URL in an environment file or directly in the notebook. That is a manageable class exercise, but credentials and generated code deserve more care in a real data workflow.
The project demonstrates a broader trait of the repository: examples are teaching units with different shapes. A Chapter 4 script, a Chapter 13 web service, and a contributed notebook do not share one deployment contract. Dependencies, model providers, data files, and external services change as the reader moves through the book. Plan to create separate environments and pin versions per exercise instead of treating the 137-package install as a course-wide environment.
Open corrections expose version drift in examples
Open issue 862 identifies a Chapter 11 comment saying batch_size=2 must be divisible by num_generations=8, even though the stated arithmetic is false. The reporter explains that effective batch size, including gradient accumulation, is the relevant quantity. This is a content correction rather than proof that the complete training routine fails, but it shows why learners should check explanations against the library version they install.
Other open pull requests propose fixes for a Chapter 12 evaluator import and result fields, a Chapter 13 health check, and response extraction in a simple agent example. Each report is specific, and several include offline regression checks. Their open status means the fixes should not be described as shipped. The useful conclusion is that fast-moving framework APIs can age faster than a long tutorial, especially when code spans 16 chapters.
An August 18 push and 189 issues and PRs show a busy queue
The default branch was last pushed on August 18, 2026. GitHub listed 189 open issues and pull requests, while issue and PR activity continued into September 2. The repository had 76,450 stars when fetched. That is an active learning community with a large correction queue; the combined open count is not a bug total, and popularity cannot substitute for runnable checks.
V1.0.3 is a PDF and content release focused on documentation fixes, code examples, formulas, images, and navigation. It is useful for identifying a stable reading snapshot, but it does not turn every contributed application into a versioned software product. Use the course as intended: read the sequence, run each example in its own environment, question stale interfaces, and add tests around any code that leaves the notebook.