Should you study it or install the whole repository?
Start with the book and one selected experiment. The chapter structure gives engineers a route through agent design, while our root-level sandbox run leaves substantial execution questions unanswered. Installation completed, but the build failed and pytest stopped during collection. That supports a narrow recommendation: evaluate each lab on its own instructions and acceptance criteria.
The English README expects readers to understand moderately complex Python, an AI coding tool, the command line, Git, JSON and REST APIs. This is a practical prerequisite check before spending time on setup. Readers who lack that background should begin with a broader introductory course.
What our run established
Our run on September 29, 2026, at commit b789c97 installed successfully, but the recorded build and test commands exited unsuccessfully.
The matching sandbox record identifies the full commit as b789c97a8f3de5b405a6f6dfa8289208d679d0b8. It used the Python 3.12 Bookworm image with 3 CPUs and 8 GB of memory. Installation took 186 seconds and recorded 338 packages with a 7,523 MB dependency footprint. These are observations from our sandbox, not estimates from the README.
Our harness stores each command's success flag, exit code, elapsed time and retained output tail, alongside parsed test counts. The build command exited 1 after 13 seconds. Its empty retained tail does not identify a cause. The test command exited 3 after 67 seconds. Neither command timed out.
The stored parser summary records 0 passed, 0 failed, 0 skipped and 42 errors. It does not establish that 42 unique tests ran or that the repository's full test coverage was examined. The raw tail shows pytest importing chapter3/agentic-rag-for-user-memory/test_enhanced_logging.py, where line 17 calls sys.exit(1). Pytest reports an unexpected SystemExit. Its footer says "42 errors in 43.64s"; that footer is separate from the harness's 67-second command duration. The visible import failure does not explain every recorded error.
Pip-audit recorded 6 advisories across 2 packages. The number of scanned packages is unknown because the record does not supply scannedPackages. These advisories neither establish overall security nor identify the reachability of an affected dependency in a selected exercise.
The MrKeyoor testing methodology explains command outcomes, sandbox limitations and advisory reporting. That page supplies provenance context; its current procedure is not evidence that this older run used every current rule. This correction uses the existing run and includes no new execution.
A successful installation establishes that the recorded install command completed. It does not establish successful model calls, browser workflows, training, robot operation or deployment. Our run did not verify those workloads.
What the book teaches, and what its code promises
AI Agent Book is an open-source book about agent design and engineering across 10 chapters.
The chapters cover agent fundamentals, context engineering, memory and knowledge bases, tools and MCP, coding agents, interaction, evaluation, post-training, continual improvement and multi-agent collaboration. The useful feature for a learner is the connection between a design question and a companion exercise. It still takes a separate check to determine whether that exercise is runnable locally, an external reproduction or a design exercise.
The README says coding agents generated all companion code and that the author does not read it line by line. It assigns the initial design and final acceptance to a human. Before reusing an example, inspect its inputs, side effects and checks, then establish whether its output meets your requirements.
The project's experiment status ledger separates retained execution evidence from source availability. It explicitly says that cloning a pinned repository, installing dependencies or passing a smoke test does not establish completion. It also distinguishes completion from a positive result: an experiment can be complete while failing to support its hypothesis.
That ledger is the project's account of its retained evidence, not work performed by our sandbox. It is selective rather than a complete numbered experiment index. Use the chapter README to find an exercise and the ledger to understand the stated evidence boundary.
Check the English instructions against the source
The supplied Chinese README lists 109 experiments; the English README lists 93. Those are documentation counts, not our test results. The project warns that community translations may lag the Chinese original.
The difference reaches beyond the headline count. The English appendix still places some external repositories under older chapter paths, while its chapter overview describes the reorganized chapters. If an English command points to a different location from the current Chinese instructions, resolve that difference before running it. Do not combine commands from the two editions without checking their intended paths.
English readers can read online or use the PDF and EPUB editions without building the book themselves. Compiling an edited manuscript has its own publishing dependencies and is separate from running an agent experiment.
Use the chapter as your setup boundary
The documented shared Python range is 3.11 through 3.13, with some vendored Chapter 8 components requiring Python 3.12 or later. The README recommends uv sync --locked --extra ch1 for the first chapter and corresponding extras for other chapters. This is documented guidance, not a command we newly tested for this correction.
The broad all extra does not include every experiment. Read the selected experiment's README for credentials and system dependencies. Some accept root-level configuration; others need a local environment file or exported variables. Local Ollama is an option only where that experiment's README or CLI lists support.
A useful next step is to choose one exercise, pin the source commit, read its acceptance requirements and install its documented dependencies in a disposable environment. Record the actual output and missing prerequisites. Keep the result scoped to that exercise rather than treating it as validation of the whole book.
Which learning resource fits your next step?
AI Agents for Beginners uses Microsoft Agent Framework and Microsoft Foundry Agent Service V2 in its code examples.
Choose Microsoft's AI Agents for Beginners as an alternative learning route if you want topic-based agent lessons and intend to work with that stack. Its README says Microsoft Foundry requires an Azure account, while some samples support alternative OpenAI-compatible providers. We have not compared its setup reliability with AI Agent Book in this run.
Generative AI for Beginners teaches the fundamentals of building generative AI applications through 21 lessons.
Choose Microsoft's Generative AI for Beginners first if you need foundations in prompts, text and chat applications, search or function calling before concentrating on agents. It can complement later study of AI Agent Book; it is not a replacement for the book's full agent-focused scope.
How to read the ratings
These are editorial judgments, not benchmark results or adoption measurements. The rubric runs from 1 to 5: 1 means a major observed obstacle, 2 means substantial limitations, 3 means mixed evidence, 4 means strong support with a material qualification, and 5 means broad support for the stated dimension within the evidence reviewed.
Setup ease is 1 because our root installation had a large footprint and the subsequent build and test commands did not succeed. Documentation is 4 because the project explains prerequisites, chapter setup and evidence status, with a material translation mismatch. Community is 5 for the breadth of documented community translations and contribution routes, not for measured popularity or support responsiveness. Maturity is 3 because the project distinguishes retained evidence from incomplete work, while our root run did not establish a working acceptance path.