Learn Agentic AI with Lavigne

Back cover body

There is no shortage of writing about AI agents. There is, however, a shortage of writing that shows you one — whole, running, and small enough to hold in your head.

Learn Agentic AI with Lavigne teaches agentic AI through a single, complete, working artifact: braagle_weasel.py, a single-file, standard-library-only skills orchestrator — the Weasel Orchestrator Agent — that runs a full agent loop, loads procedural knowledge on demand, records every decision it makes, and tests itself offline.

This is not a book about imaginary architecture. It is a guided walk through a working system: skills, tools, state, stop logic, loop mechanics, tracing, regression, and the boundary between model judgment and guarantees.

By the end, the agent is no longer magic. It is mechanics you can read, run, trace, and understand.

Inside you’ll learn

• how skills package procedural judgment
• how tools expose capabilities safely
• why state needs one writer
• how stop logic keeps the loop honest
• how traces become regression tests
• where model judgment ends and guarantees begin

Grounding statement

This book is grounded in a single running artifact: braagle_weasel.py, version 1.0.0, an 838-line, single-file, standard-library-only Weasel Orchestrator Agent with a passing self-test. Every technical claim in the book traces to that artifact, its teaching primer, or its complete operating manual, the context pack.

The book does not describe an imagined framework, a diagram-only architecture, or a vendor abstraction. It explains the actual code that runs: the skill loader, tool contract, state chokepoint, stop logic, loop mechanics, trace emitter, regression harness, mock model, fixture skills, business tools, and self-test.

Where the book expands beyond the source, it does so pedagogically, not inventively. Analogies, walkthroughs, exercises, and chapter compressions are used to make the mechanics legible, but they do not add capabilities the framework lacks.

The result is a book about agentic AI with a narrow promise: no hidden machinery, no ungrounded claims, and no component that exists only in prose.