Alex Chen

The open-source repository bojieli/ai-agent-book contains the Chinese book Understanding AI Agents: Design Principles and Engineering Practice, a compiled PDF, and chapter-based code. It uses the Apache-2.0 license.

A fast-growing star count can tell you that many people noticed a repository. It cannot tell you that you learned the material.

Here is a simple way to use the book as a lab manual instead of another saved link.

Make an evidence sheet

For every chapter you study, create one record:

chapter: 1
concept: tool-using agent loop
example_path: <repository path>
command: <exact command>
expected_output: <one observable result>
actual_output: <fill after running>
failure_changed: <one input you changed>
what_i_learned: <one sentence>

Enter fullscreen mode Exit fullscreen mode

Do not mark a chapter complete because you read it. Mark it complete when you can run one example, predict one output, and explain one failure.

The three-pass method

Pass 1: reproduce

Pin a commit, create the environment described by the repository, and run the smallest example for the chapter. Record Python version, dependencies, command, and output.

Pass 2: break

Change one bounded input. For an agent loop, remove a tool, return malformed JSON, or reduce the step limit to one. Predict the failure before running it.

Pass 3: explain

Write the lifecycle without framework names:

input -> model decision -> validation -> tool execution -> observation -> stop

Enter fullscreen mode Exit fullscreen mode

If you cannot identify where the failed input was rejected—or should have been rejected—repeat the experiment.

A small progress test

After three chapters, rebuild one example without copying the framework glue. A useful beginner target is a deterministic two-tool loop:

  • add(a, b) accepts two numbers;
  • weather(city) returns a fixed fixture;
  • invalid tool names are rejected;
  • the loop stops after three steps.

Expected evidence includes one successful call, one schema failure, and one step-limit failure.

Common mistakes

  • Installing everything at once: use the smallest chapter environment first.
  • Treating reported results as your results: keep source evidence and local evidence in separate fields.
  • Skipping failures: the failure is often where the architecture becomes understandable.
  • Chasing completion percentage: one reproduced concept is more valuable than ten skimmed chapters.

Limitations

I have not executed every example in the repository. Its contents and setup instructions can change, so pin the commit you study and follow the current README. This workflow measures your reproducibility and understanding; it does not rate the book's completeness.

A star can put a book on your radar. Only an experiment puts the concept in your hands.