Build an AI Agent (From Scratch)

Build an AI Agent (From Scratch), by Jungjun Hur and Younghee Song, was published by Manning in July 2026 (336 pages). It builds one Python research agent across ten chapters and deliberately avoids agent frameworks, so every layer is code you wrote. Part 1 covers the foundations: what separates a fixed workflow from an LLM-directed agent loop, choosing models and calling stateless LLM APIs, prompt engineering, tool definitions and MCP, and then a full ReAct agent with an ExecutionContext store, a tool abstraction, an LLM communication layer and structured output, all scored on GAIA tasks. Part 2 adds the advanced capabilities: vector and structure-based RAG with callbacks, agent memory (in-run context management, sessions and state, long-term memory across sessions), planning and reflection, code execution in cloud sandboxes with a filesystem workspace and agent skills, multi-agent orchestration (workflows, agent-as-tool, agent transfer, and A2A across networks), and an evaluation chapter on observability, dataset building, LLM-as-a-judge grading and CI/CD. The MIT-licensed companion repo (shangrilar/ai-agent-from-scratch) packages the finished agent as an installable scratch_agents library, uses uv and Python 3.13, and includes a pytest suite that runs without API keys. Hur is an independent AI and data engineer who builds agentic systems for e-commerce and wrote Practical AI Application Development Using LLMs; Song is an AI consultant at PwC.

Topic

Agentic

Level

Intermediate

Format

Book

Cost

Paid

Duration

336 pages, 10 chapters plus runnable notebooks; roughly 25-35 hours self-paced (estimate)

Provider

Manning Publications (Jungjun Hur and Younghee Song)

Hands-on

Yes — code/exercises

Certificate

None

• ✓Tell a fixed developer-defined workflow apart from an LLM-directed agent loop, and know when each fits
• ✓Define tools for LLM function calling and standardize them through the Model Context Protocol
• ✓Implement a ReAct agent with an execution context, tool abstraction and an LLM communication layer
• ✓Measure your agent's real limits by running it against tasks from the GAIA benchmark
• ✓Add vector search and structure-based retrieval to an agent, and extend it with callbacks
• ✓Manage in-run context, session state and long-term memory that persists across agent sessions
• ✓Use explicit planning and reflection steps to reduce looping and hallucination on multi-step tasks
• ✓Give an agent sandboxed code execution, a filesystem workspace, CLI access and reusable agent skills
• ✓Orchestrate multiple agents with workflows, agent-as-tool, agent transfer and the A2A protocol
• ✓Evaluate agents with observability, curated datasets, LLM-as-a-judge grading and CI/CD checks

• •No agent framework anywhere: the final scratch_agents package is code you built yourself, so you can see what LangGraph-style libraries hide
• •GAIA benchmark tasks run through the whole book as a fixed test, so you can see each new capability move the score
• •Current agent topics you rarely find in one book: MCP, A2A, agent skills, cloud sandboxes and context rot
• •The companion repo is MIT-licensed, uses uv and Python 3.13, ships per-chapter snapshot files and has a test suite that runs without API keys
• •Ends on evaluation and CI/CD, the part most agent tutorials skip

This page was written from 3 sources, 1 on domains other than manning.com.

1. 1.manning.com — build an ai agent from scratchvendor
2. 2.manning.com — chapter 1vendor
3. 3.github.com — ai agent from scratch