Where Is the Shared Plumbing Code for AI Agent Experiments in ai-agent-book?

TLDR: In the bojieli/ai-agent-book repository, the shared plumbing code for AI agent experiments lives in the agentbook package, primarily in agentbook/__init__.py, which centralizes provider resolution, environment loading, and trace printing across all experiments.

The bojieli/ai-agent-book repository organizes a series of AI-agent experiments into chapters, each with its own code and dependencies. But when you dig into the structure, you'll find that a single, minimal package does much of the heavy lifting behind the scenes: agentbook. This shared plumbing handles the common tasks every experiment needs — resolving providers, loading environment variables, and printing traces — so each chapter stays focused on its specific logic. In this guide, we'll walk through exactly where this shared plumbing lives, how it works, and which files matter most.

What Is the Shared Plumbing Code in Agentbook?

What Is the Shared Plumbing Code?

The shared plumbing in this repository is a marked package dedicated only to infrastructure, not experiment logic. Both concept and content are described at the top of agentbook/__init__.py.

According to the source file, the package contains:

"plumbing only – provider resolution, environment loading, trace printing."

Just line in agentbook/__init__.py also defines the __version__ metadata. It's these small, foundational pieces that every chapter's code imports when experimenting, so keeping them centralized means the experiments don't repeat the same provider setup, config reading, or trace output logic.

Inside this shared layer, the concrete provider logic is pushed into the submodules under agentbook/providers/. With the package acting as an entry point, all the "how" — the actual provider registry, network calls, and model lookups — is kept organized and testable.

Locating the Shared Plumbing: key Package Structure

The repository's experiments concord in chapters/, and bonus code in bonus/, but they all access the same agentbook package. Here's a quick overview of the physical structure:

ai-agent-book/
├── agentbook/
│   ├── __init__.py          # plumbing entry point, version

│   ├── providers/           # all provider-resolution logic

│   │   ├── registry.py      # register/retrieve providers

│   │   ├── openrouter.py    # OpenRouter integration

│   │   ├── models.py        # model spec definitions

│   │   └── legacy.py        # backward-compatible provider layer

└── chapters/                # per-chapter agent experiments

agentbook acts as that shared "plumbing" layer for repository-wide avoidance of duplication. While each chapter may use a handful of functions, they all repel the same underlying registry and env loading helpers.

Exploring the agentbook.providers Subpackage

The real logic of the shared plumbing is in agentbook/providers/:

registry.py — This is where providers are registered and retrieved. It's the entry point for any experiment wanting to select a model, e.g., by name. A typical call looks like:

from agentbook.providers import registry

model = registry.get_provider("gpt-4o-mini")
print(model)  # Provider instance backed by the registry

**openrouter.py

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