跪拜 Guibai
← All articles
Artificial Intelligence · Interview · GitHub

A Full AI Agent Engineering Book, with 28 Runnable Projects, Is Now Open-Source

By Cosolar ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most agent resources stop at API tutorials. This book supplies the missing engineering layer — context design, tool sandboxing, small-model validation, and full evaluation pipelines — that turns a demo into a production system.

Summary

The book, now fully open-sourced with its PDF and code, builds everything around the equation Agent = LLM + Context + Tools. It spans from context engineering and KV Cache-friendly design to production-grade Coding Agents, multi-agent collaboration, and self-evolving systems. Each of the 10 chapters includes runnable projects, such as a 17-tool coding assistant, a hybrid retrieval pipeline for Agentic RAG, and a dual-agent phone-and-browser parallel system.

A deliberate design choice runs throughout: small models are used to prove that system architecture matters more than raw scale. Chapter 2, for instance, demonstrates complete tool calling on a 0.6B local model. The projects are ablation-driven, with controlled experiments that quantify the impact of removing individual context components, and the Coding Agent is built in pure Python with no command-line dependencies.

Li Bojie, a former Huawei Fellow and current agent infrastructure founder, targets engineers who need to build real systems, not just call APIs. The book also covers evaluation benchmarks like SWE-bench and GAIA, post-training comparisons between SFT and RL, and practical API key guidance for Chinese and international model providers.

Takeaways
A 0.6B local model can perform complete tool calling when the surrounding system design is sound, making scale a secondary concern.
Context engineering, not prompt tweaking, is the primary lever for agent performance; ablation experiments quantify the effect of each component.
The Coding Agent runs in pure Python with no CLI dependencies, and its MCP servers include LLM secondary approval and sandbox isolation.
Self-evolution is implemented as a closed loop: experience learning, workflow recording into toolification, and proactive tool discovery.
Evaluation covers Terminal-Bench, SWE-bench, and GAIA, plus a cost breakdown and TTS quality assessment, giving a full picture of agent capability.
Conclusions

The book's central formula — Agent = LLM + Context + Tools — is a deliberate pushback against the idea that better models alone solve agent problems; it treats the LLM as just one component in a larger system.

Validating tool calling on a 0.6B model is a practical counter to the assumption that agent capabilities require frontier models, which has direct cost implications for anyone running agents at scale.

Open-sourcing a full textbook with 28 runnable projects, rather than just a code repo, signals a bet that agent engineering is now a distinct discipline with its own canon, not a subfield of prompt engineering.

Concepts & terms
Agent = LLM + Context + Tools
A design formula that frames an AI agent not as a standalone model, but as a system composed of a large language model, its surrounding context (memory, instructions, conversation history), and the tools it can invoke.
KV Cache-friendly design
A context engineering approach that structures prompts and conversation history to maximize reuse of the key-value cache inside a transformer model, reducing latency and compute cost on long interactions.
Agentic RAG
Retrieval-Augmented Generation where the agent actively decides when and how to retrieve information, rather than following a fixed retrieval step, enabling multi-step reasoning over external knowledge.
MCP (Model Context Protocol) servers
Servers that expose tools to an agent via a standardized protocol, categorized here into perception, execution, and collaboration types, with optional LLM-based approval and sandboxing for safety.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗