What Is the Specific Structure of Agent Workbench Lessons? Inside the ai-engineering-from-scratch Curriculum

Every Agent Workbench lesson in the rohitg00/ai-engineering-from-scratch repository is a self-contained directory under phases/14-agent-engineering/ that packages a documented tutorial, a mission statement, executable stdlib Python code, at least five unit tests, a six-question quiz, and optional reusable artifacts.

The rohitg00/ai-engineering-from-scratch repository teaches agent engineering through a strict, reproducible lesson format. The Agent Workbench lessons—located in Phase 14 · Agent Engineering—follow a standardized directory and file contract that supports the curriculum’s seven workbench surfaces: Instructions, State, Scope, Feedback, Verification, Review, and Handoff. This layout guarantees that each lesson can be built, tested, and shipped independently.

Standard Directory Layout for Every Agent Workbench Lesson

Each lesson lives in its own folder under phases/14-agent-engineering/<lesson-slug>/ and contains the same core set of files. For example, the lesson Agent Workbench: Why Capable Models Still Fail resides in phases/14-agent-engineering/31-agent-workbench-why-models-fail/.

The complete tree looks like this:

phases/14-agent-engineering/<lesson-slug>/
├─ docs/
│   └─ en.md                 # tutorial + front-matter

├─ mission.md                # one-sentence mission

├─ code/
│   ├─ main.py               # implementation (Python stdlib)

│   └─ tests/
│       └─ test_main.py      # ≥5 unit tests

├─ quiz.json                 # 6-question knowledge check

└─ outputs/                  # optional reusable artifacts

Core Components of the Agent Workbench Lesson Structure

Lesson Documentation (docs/en.md)

The file docs/en.md is the human-readable tutorial. It includes front-matter that satisfies the curriculum contract defined in AGENTS.md, ensuring the lesson metadata aligns with the README.md table and the ROADMAP.md status row. The metadata specifies the title, type, languages, prerequisites, time, and learning objectives.

For example, phases/14-agent-engineering/31-agent-workbench-why-models-fail/docs/en.md opens with:

# Agent Workbench Engineering: Why Capable Models Still Fail

> A capable model is not enough. Reliable agents need a workbench…

**Type:** Learn + Build  
**Languages:** Python (stdlib)  
**Prerequisites:** Phase 14 · 01 (Agent Loop), Phase 14 · 26 (Failure Modes)  
**Time:** ~45 minutes

Mission Statement (mission.md)

The mission.md file sits at the lesson root and provides a concise, one-sentence mission description. This string appears in the UI and helps learners orient themselves before they start the tutorial.

Executable Code (code/main.py)

Every lesson includes a pure stdlib Python script at code/main.py. The script demonstrates the workbench surface in action and can be executed directly without external dependencies. In 31-agent-workbench-why-models-fail/code/main.py, the implementation runs a minimal repo-task twice—once as a prompt-only run and once with the full workbench surfaces wired in—and writes a failure_modes.json report.

Unit Tests (code/tests/test_main.py)

Lessons validate correctness with a language-native test suite located at code/tests/test_main.py. The repository requires at least five unit tests per lesson. You can execute them with the stdlib runner:

cd phases/14-agent-engineering/31-agent-workbench-why-models-fail/code
python3 -m unittest discover tests -v

Knowledge Quiz (quiz.json)

At the root of each lesson, quiz.json encodes a six-question assessment. The format follows a strict split: one pre-question, three checkpoint questions, and two post-questions. This enforces the learning outcomes after the student has read the documentation and run the code.

Reusable Outputs (outputs/)

The outputs/ directory is optional, but when present it holds reusable artifacts that downstream lessons can import. For instance, phases/14-agent-engineering/31-agent-workbench-why-models-fail/outputs/skill-workbench-audit.md is a skill file generated for later reuse. The capstone lesson at phases/14-agent-engineering/42-agent-workbench-capstone/outputs/agent-workbench-pack/README.md even ships a full pack description.

How the Seven Surfaces Map to the Lesson Code

The Agent Workbench curriculum is built around seven surfaces: Instructions, State, Scope, Feedback, Verification, Review, and Handoff. The code/main.py script in each lesson is the concrete implementation of these surfaces.

In the Why Capable Models Still Fail lesson, the script contrasts a naive prompt-only execution against a run that uses the full workbench. It then persists its findings to failure_modes.json, turning the abstract surface concepts into observable behavior that the unit tests assert against.

Running an Agent Workbench Lesson Locally

Because every lesson relies only on the Python standard library, you can run them without installing extra packages. Navigate to the lesson’s code/ directory and execute the script:

cd phases/14-agent-engineering/31-agent-workbench-why-models-fail/code
python3 main.py

Then validate your changes with the bundled test suite:

python3 -m unittest discover tests -v

This "one commit per lesson" workflow guarantees that each Agent Workbench lesson remains self-contained, reproducible, and independently shippable.

Summary

• Every Agent Workbench lesson lives under phases/14-agent-engineering/<lesson-slug>/ in the rohitg00/ai-engineering-from-scratch repository.
• The standardized structure includes docs/en.md, mission.md, code/main.py, code/tests/test_main.py, quiz.json, and an optional outputs/ folder.
• docs/en.md carries curriculum-compliant front-matter that links the lesson to the README.md table and ROADMAP.md status row.
• code/main.py demonstrates the seven workbench surfaces with pure stdlib Python and emits artifacts such as failure_modes.json.
• Each lesson ships with at least five unit tests and a six-question quiz to enforce learning outcomes.
• The capstone lesson produces a reusable agent-workbench-pack that includes its own README.md.

Frequently Asked Questions

Where are the Agent Workbench lessons located in the repository?

The Agent Workbench lessons are part of Phase 14 · Agent Engineering and are stored under the phases/14-agent-engineering/ directory. Each lesson has its own subfolder, such as phases/14-agent-engineering/31-agent-workbench-why-models-fail/.

What is the purpose of the mission.md file?

The mission.md file contains a concise, one-sentence mission statement that surfaces in the UI. It orients the learner by summarizing the lesson’s goal before they open the tutorial or the code.

How many questions are in an Agent Workbench lesson quiz?

Every quiz.json contains exactly six questions: one pre-assessment question, three checkpoint questions, and two post-assessment questions. This structure validates knowledge before, during, and after the lesson.

Do Agent Workbench lessons require third-party Python libraries?

No. According to the source code, each lesson’s code/main.py is written using the Python standard library only. Learners can run the scripts directly with python3 main.py and execute tests with python3 -m unittest discover tests -v without installing external dependencies.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}