1. Objectives

After studying this manual, the reader should understand:

• What an LLM-based agent is and what it is not.
• The distinction between the LLM engine and the agent harness.
• How the model influences agent performance.
• How the runtime harness influences reliability, safety, cost, and diagnosability.
• The core layered architecture of a simple agent.
• How to design a simple agent from first principles: context, tools, loop, guardrails, memory, and observability.
• Common failure modes and how to map them to architectural fixes.