Type: url
Original: https://jimmysong.io/zh/book/ai-handbook/agent/engineering/
Status: partial
Summary
A draft chapter of Jimmy Song’s online handbook 智能体构建指南 arguing that LLM-based agents must move from merely running to being reproducible, extensible, observable and maintainable through systems engineering rather than prompt tweaking. It describes three layers of complexity (runnability, reproducibility, evolvability), shows how per-call uncertainty compounds across many LLM calls (90% per call gives about 35% over 10 calls and 12% over 20), and prescribes replay, version tracking, monitoring metrics and checkpoint-based error recovery. It closes with a five-stage growth model, a table of agent design patterns and a layered engineering roadmap, concluding that an agent’s future lies in being more reliable rather than smarter.