"AI 에이전트 엔지니어링" (AI Agent Engineering), authored by Michael Albada and published by Hanbit Media, serves as an all-in-one guide for developing AI applications, ranging from single-agent to multi-agent systems. The book addresses the growing need for enterprises to leverage AI agents—systems that combine tools, knowledge, and memory to solve complex problems beyond simple automation—while acknowledging the inherent challenges in their design, orchestration, and deployment. It aims to provide practical, hands-on guidance with clear design principles to facilitate the rapid conversion of ideas into production-ready solutions.

The book is structured into three main parts, systematically covering the lifecycle of AI agent systems:

Part 1: Agent System Concepts and UX
This section lays the foundational understanding of AI agents and their user experience.

• Chapter 1: Agents defines AI agents, discusses the transformative impact of pre-trained foundation models, categorizes agent types, and explores model selection. It covers the shift from synchronous to asynchronous interactions, outlines various use cases, and delves into workflow integration. Crucially, it introduces core principles for building effective agentic systems, organizational strategies for their construction, and an overview of agentic frameworks.
• Chapter 2: Agent System Design details the architectural components and considerations for building agent systems.
  • Core Components: The book elaborates on the importance of Model Selection (likely referring to the choice of Large Language Models or other foundation models). Tools are a central concept, emphasizing the design of specific tools for tasks (e.g., API calls, external functions) and their modular integration. Memory is broken down into short-term memory (managing the context window, $C_w$) and long-term memory (e.g., semantic memory, vector stores, and Retrieval Augmented Generation (RAG)). This involves concepts like embedding generation, similarity search for information retrieval, and knowledge graph integration for structured information. Orchestration is the mechanism that controls the agent's decision-making process and sequencing of actions, guiding it to select and utilize tools and memory effectively.
  • Design Trade-offs: The text highlights critical engineering considerations such as balancing Performance (speed vs. accuracy), ensuring Scalability for handling increased loads, maintaining Reliability (robustness and consistency), and managing Cost-effectiveness.
  • Architecture Design Patterns: It distinguishes between Single Agent Architectures and Multi-Agent Architectures, focusing on collaboration, parallelism, and coordination among agents.
  • Best Practices: Emphasizes incremental design, robust evaluation strategies, and thorough real-world testing.
• Chapter 3: UX Design for Agent Systems focuses on user interaction, covering different modalities (text, graphical, voice, video, and their combinations), synchronous vs. asynchronous agent experiences, context maintenance, continuity, and clear communication of agent capabilities, including uncertainty and failure handling. Trust building in interaction design is also discussed.

Part 2: Building and Scaling Agent Systems
This part provides practical methodologies for implementing and expanding agent systems.

• Chapter 4: Tools dives into tool implementation, using LangChain as a primary example. It covers local, API-based, and plugin tools, along with concepts like Multi-hop Code Path (MCP) and stateful tools. It also touches upon automating tool development using foundation models and real-time code generation.
• Chapter 5: Orchestration explores various agent types, including reflective agents, ReAct agents (Reasoning and Acting), plan-and-execute agents, query decomposition agents, introspective agents, and deep research agents. It details Tool Selection mechanisms (standard, semantic, hierarchical) and Tool Execution (single, parallel, chained, graph-based). The concept of Context Engineering is also a key focus, emphasizing the careful construction of prompts and information provided to the agent to guide its behavior and decision-making.
• Chapter 6: Knowledge and Memory elaborates on memory management, covering context window management and full-text search. It details Semantic Memory and Vector Stores, including semantic search, implementation with vector databases, and the crucial RAG (Retrieval Augmented Generation) paradigm for grounding LLM responses with external knowledge. GraphRAG, leveraging knowledge graphs for enhanced retrieval, and dynamic knowledge graphs are also introduced.
• Chapter 7: Learning of Agentic Systems distinguishes between Non-parametric Learning (e.g., learning from examples, reflection, experiential learning) and Parametric Learning (Fine-tuning). The latter includes fine-tuning large foundation models, small models, Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning from Human Feedback (RLHF) for verifiable rewards.
• Chapter 8: From Single Agents to Multi-Agents discusses the rationale and principles for expanding to multi-agent systems, including "swarms." It explores Multi-Agent Coordination strategies such as democratic, supervisor-centric, hierarchical, and actor-critic approaches. Techniques for automated agent system design, inter-agent communication protocols (A2A), message brokers, event buses, actor frameworks, orchestration/workflow engines, and state/persistence management are also covered.

Part 3: Operating and Governing Reliable Agents
This section focuses on the operational aspects, ensuring trustworthiness, security, and human-agent collaboration.

• Chapter 9: Verification and Measurement emphasizes monitoring agentic systems, integrating evaluation into the development lifecycle, and expanding evaluation datasets. It covers component evaluation (tools, planning, memory, learning) and holistic evaluation (end-to-end performance, consistency, coherence, hallucination detection, handling unexpected inputs).
• Chapter 10: Operational Environment Monitoring details monitoring stacks (Grafana, ELK, Arize Phoenix, Signoz, Langfuse) and the importance of OpenTelemetry for instrumentation. It discusses visualization, alerting, and monitoring patterns like shadow mode, canary deployments, regression trace collection, and self-healing agents, alongside user feedback integration and distribution shift detection.
• Chapter 11: Improvement Loops describes continuous improvement through feedback pipelines, including automated issue detection, root cause analysis, human intervention for review, prompt/tool refinement, and prioritizing improvements. It covers experimentation (shadow deployment, A/B testing, Bayesian bandit) and continuous learning (in-context learning, offline retraining).
• Chapter 12: Security of Agentic Systems addresses unique risks, new attack vectors, and Foundation Model Security (defense techniques, red teaming, MAESTRO threat modeling). It focuses on Data Protection (privacy, encryption, provenance, integrity, sensitive data handling) and Agent Security (guardrails, protection from external threats, internal failures).
• Chapter 13: Human-Agent Collaboration explores the roles and autonomy in agent systems, stakeholder alignment, and driving adoption. It covers expanding collaboration through agent scope, organizational roles, shared memory, and context boundaries. The chapter concludes with discussions on Trust, Governance, and Compliance (trust lifecycle, accountability frameworks, response procedures, privacy, and regulatory compliance), envisioning the future of human-agent teams.

The book is highly recommended for machine learning/software engineers looking to deploy robust prototypes to production, backend/platform engineers designing multi-agent orchestrations, and technical product managers/owners seeking to integrate agents into workflows with comprehensive operational frameworks. It stands out by providing a balanced view of academic trends and practical methodologies, covering the entire lifecycle of AI agent systems from design and implementation to operations, security, and governance, going beyond mere concept explanations or framework usage guides.