AI Agent Book: a study map for agent engineering
Start with the context-ablation experiment
Run one documented chapter-1 example and keep its provider and mode visible.
What you will learn
- Prepare a bounded environment
- Change one context component
- Keep setup and outcome separate
Before you start
- Python environment basics
- Access to a supported model provider for live experiments
Turn one context lesson into a documented experiment and design decision.
Key takeaways
- Use one experiment before installing the full course.
- Change only one controlled variable initially.
- Generated files are not proof of valid inference.
Prepare a bounded environment
The root README recommends a locked uv environment, with pip as a fallback. The chapter guide documents provider credentials and a context experiment under chapter1/context. Use the repository lock and follow the local instructions for the experiment you chose.
Keep a provider key in a secret file or environment, never in an article or output attachment. Choose a small task without sensitive documents.
Change one context component
The chapter lists full, no_history, no_reasoning, no_tool_calls and no_tool_results modes. Run the same task with two modes and save raw output, provider, model and revision.
A result that changes between runs does not by itself prove the removed component caused it. Repeat trials and inspect the actual request messages.
Keep setup and outcome separate
The runner can write evidence files, but an output directory existing is not a successful model experiment. Inspect completion, correctness and whether the intended context was removed.
This article does not claim to have executed the example. The command below is a reading cue; confirm flags in the pinned experiment guide before use.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Set up the locked environment and one provider key.
- 2
Run a small synthetic task in full and no_tool_results modes.
- 3
Save requests, outcomes and an explanation of the difference.
Copy-ready example
uv sync
uv run python chapter1/context/main.py
# inspect chapter1/context/README.md for provider and mode flagsFrequently asked questions
Must I use the default provider?
The chapter guide documents several providers; select one supported in your environment.
Is no_reasoning a guarantee that the model never reasoned?
No. The mode concerns retained context, and provider behavior may differ.
Sources
- AI Agent Book / README.mdSource checked 2026-09-26
- AI Agent Book / chapter1/context/README.mdSource checked 2026-09-26