Volume 8 — AI Agents¶

Applied AI Engineering | Volume 8 of 10

Introduction¶

The shift from static LLM completions to autonomous, tool-using AI agents is one of the most consequential transitions in applied AI engineering. An AI agent does not merely respond — it perceives its environment, reasons about goals, selects and invokes tools, and iterates until a task is complete or a stopping criterion is met. This loop unlocks capabilities that a single prompt-response cycle cannot: searching the web, writing and executing code, querying databases, coordinating with other agents, and adapting its own plan mid-execution.

This volume provides a rigorous, hands-on treatment of AI agents. It starts from first principles — what an agent is, how its components fit together, and why the ReAct pattern works — moves through the dominant frameworks used in industry today, and concludes with the complex challenges of multi-agent systems: coordination, communication, consensus, and safety.

By the end of this volume you will be able to design, build, evaluate, and operate agentic AI systems that go well beyond simple chatbots.

Chapters¶

Chapter Dependency Graph¶

graph LR
    PRE["Volume 7 — RAG"] --> C1
    C1["Ch 1 · Agent Fundamentals<br/>(ReAct, tools, memory)"] --> C2
    C1 --> C3
    C2["Ch 2 · Agent Frameworks<br/>(LangChain, LlamaIndex, SDK)"] --> C3
    C3["Ch 3 · Multi-Agent Systems<br/>(Orchestration, security)"]

    style PRE fill:#64748b,color:#fff
    style C1 fill:#7c3aed,color:#fff
    style C2 fill:#7c3aed,color:#fff
    style C3 fill:#7c3aed,color:#fff

Chapter 1 is a prerequisite for both Chapter 2 and Chapter 3. Readers already comfortable with the ReAct pattern may skip to Chapter 2 but should review the memory and evaluation sections first.

Learning Outcomes¶

On completing this volume you will be able to:

1. Explain the Sense–Think–Act loop and identify each component in a real agentic system.
2. Implement a ReAct agent from scratch using Anthropic's tool_use API, with correct loop termination conditions.
3. Select the right agent framework (LangChain, LlamaIndex, CrewAI, or the Anthropic Agents SDK) for a given production use case based on maturity, abstraction level, and observability needs.
4. Design a multi-agent architecture — choosing between hierarchical, peer-to-peer, and orchestrator-worker topologies — and justify the trade-offs.
5. Evaluate agent systems using task-completion rate, steps-to-completion, and tool-call accuracy, and identify common failure modes before they reach production.
6. Apply security best practices to isolate agents and prevent prompt-injection attacks across agent boundaries.