Early access · v0.9 · updated quarterly
Everyone has seen the demo. This book is about what comes after.
Building Agentic Systems — A Practical Guide to AI System Engineering · Yuri Syuganov
┌─────────────────┐ ┌──▶│ reason │ │ └────────┬────────┘ │ ▼ │ ┌─────────────────┐ │ │ act │ │ └────────┬────────┘ │ ▼ │ ┌─────────────────┐ └───│ observe │ └─────────────────┘
Agents fail in ways traditional software doesn’t: silently, expensively, and with great confidence. Building Agentic Systems is an engineer’s guide to closing the gap between a promising prototype and a system you can trust with real work.
- The agent loop, tool contracts, and structured outputs that make model behavior testable
- State, memory, and context budgets that survive long-running work
- Multi-agent orchestration without multiplying failure modes
- Reliability, cost, observability, and security — the production disciplines, adapted
- The operator practice: running AI-assisted projects day to day, with humans in the loop
26 chapters · 437 pages in print
What’s inside
Foundations: What Agents Actually Are
Past the “autonomous agent” myth to what you’re actually building — and the reason–act–observe loop underneath every agent that works.
Core Building Blocks
Structured outputs and contracts, tools and function calling, execution models, state, and memory — through your first end-to-end agent.
Multi-Agent Systems
From one agent to many: orchestration patterns, parallel execution and conflict resolution, and human-in-the-loop design that isn’t a rubber stamp.
Making It Production-Grade
Reliability engineering, observability and debugging, cost engineering, evaluation and testing, security and governance.
Advanced Patterns and Real Systems
Long-running resumable workflows, repo-aware agents, the AI project operator, and designing agent systems for teams.
Context and Outlook
Frameworks — when to use them and when not to — and the evolving landscape.
A living book
The field this book covers moves quarterly, and the book moves with it. Model names, prices, and context windows are re-checked against provider documentation at every release — and every update is free for as long as the book exists. You buy the book once; it stays current.
All twenty-six chapters are written — this is a complete draft, not a partial one. What stands between 0.9 and 1.0 is a final verification pass over the fastest-moving facts. Early-access readers get 1.0 and every quarterly refresh after it, free.
next v1.0 · verification pass
Built with the workflow it teaches
This book is written, reviewed, and maintained with the agentic system described in its own chapters: a human operator setting direction and reviewing every change, AI agents doing delegated work under version control. When the mid-2026 model lineup replaced the one the examples were first written against, parallel review agents swept all twenty-six chapters against the live API reference — and a human reviewed the diff. The machinery in Part V produced the pages you’ll be reading.