Build an AI Agent (From Scratch) - MEAP Version 3 (Jungjun Hur, Younghee Song)（Z-Library）

# Build an AI Agent (From Scratch) - Reading Guide ## 【One-Line Pitch】 A hands-on, framework-free guide to building LLM-powered AI agents from the ground up—perfect for Python developers who want to truly understand how agents work under the hood rather than relying on black-box abstractions. ## 【Book Arc】 - **Opening (~0%–9%)**: Introduces the agent landscape, defines what makes an agent (LLM + tools + loop), and explains why building from scratch matters for debugging and troubleshooting real-world agent failures. - **Early (~9%–28%)**: Covers LLM fundamentals, the seven levels of agency, and the critical distinction between workflows (developer-defined flows) and agents (LLM-directed flows), including patterns like single calls, chains, and routers. - **Early–Middle (~28%–47%)**: Explores when agents are actually necessary (task complexity, value, error cost), introduces the GAIA benchmark as a training ground, and establishes context engineering as the discipline that determines agent success. - **Middle (~47%–53%)**: Begins the practical implementation journey, setting up the core ReAct loop pattern (Thought → Action → Observation) that will be built upon throughout the book. - **Late (~53%–100%)**: Progresses through advanced capabilities—tool integration via MCP, agentic RAG, memory modules, planning and reflection, code execution agents, and multi-agent coordination—before covering production deployment concerns like evaluation and monitoring. ## 【Key Takeaways】 - **Agents = LLM + tools + loop** (Early): The core definition is simple—an LLM brain that decides, external tools that act, and a loop that repeats until the goal is achieved. This mental model underpins everything else in the book. - **Workflows vs. agents is a spectrum, not a binary** (Early): Seven levels of agency exist between fully developer-controlled flows and fully autonomous agents. The pragmatic question isn't "which approach?" but "where does agent behavior justify its cost?" - **Building from scratch is about debugging** (Opening): Framework lock-in hides failure points. Understanding each component internally—whether the LLM misread context or a tool returned unexpected results—is essential for diagnosing agent failures. - **Not every task needs an LLM, and not every LLM task needs an agent** (Middle): Evaluate task complexity, value, and error cost before choosing an approach. Simple, predictable tasks are better served by deterministic code or workflows. - **GAIA provides the ideal training ground** (Middle): The benchmark's clear answers enable fast feedback loops, and its web-search-heavy problems require no domain expertise—making it perfect for practicing the observe-analyze-improve cycle. - **Context engineering, not prompting, determines success** (Middle): The distinction between prompts (user input) and context (everything the LLM receives) is fundamental. How you engineer that context shapes whether an agent succeeds or fails. - **Agent development is an iterative cycle** (Middle): The core practice is identifying failures and reducing them through repeated cycles of observation, analysis, and improvement—not one-shot perfection. ## 【Reading Tips】 - **Deep-read Chapter 1** (~0%–34%): The workflow vs. agent distinction and the seven levels of agency are conceptual foundations you'll reference throughout. Don't skim this—it shapes all architectural decisions later. - **Skim the agent landscape survey** (~4%–9%): The overview of personal, customer-facing, and specialized agents is useful context but not critical to implementation. Move quickly through this section. - **Pay special attention to the GAIA discussion** (~44%–47%): Understanding why the authors chose this benchmark and how they use it to measure progress will help you track your own learning as you build. - **Treat the decision frameworks as checklists** (~34%–44%): The criteria for "does this need an LLM?" and "does this need an agent?" are practical tools you'll reuse in your own projects. Consider writing them down. - **Note that this is a MEAP (early access) version**: The excerpts cover primarily the foundational chapters. Later chapters on MCP integration, memory, planning, and multi-agent systems are mentioned but not detailed in this sample. ## 【Coverage Limits】 This guide is based on excerpts covering roughly the first half of the book (through Chapter 1 and into early implementation). Detailed content on MCP tool integration, agentic RAG, memory modules, planning/reflection, code agents, multi-agent systems, and production deployment is not covered in the available material. ##