A Three-Year Retrospective on AI Agent Architecture Evolution
Programmer Yupi’s Juejin retrospective organizes AI agent development into a chronological sequence, tracing ideas such as ReAct, tool use, agent loops, subagents, agent teams, harnesses, and super-agents. The author frames each architecture as addressing limitations of earlier approaches, from one-turn chatbots to systems designed to coordinate multiple tasks and remain active across channels.
The article offers a single explanatory framework for understanding the rapidly expanding terminology around AI agents, although its historical progression and broad capability claims are not independently benchmarked.
大家好,我是程序员鱼皮。
还记得 23 年初,我刚从鹅厂离职创业那会儿,AI 在国内还是「小荷才露尖尖角」。
那时候的 AI 只会陪你聊天,你问一句,它答一句。
没想到 3 年后的今天,AI 已经能自己开发程序、作图办公、剪辑视频,还能作为你的 24 小时全能助理,帮你盯群消息、回邮件、追热点,甚至在你睡觉的时候发现 Bug 并主动修好。
这 3 年里,AI 圈儿冒出了一大堆新词,什么 ReAct、Agent Loop、Subagent、Agent Teams、Harness、Loop Engineering、Personal Agent、Super-agent 等等,很多朋友直呼学不动了。

Some source content is omitted. Continue reading on the original site.
#AutoGPT#Claude Code#DeepSeek Harness#Anthropic Agent Skills#OpenAI Function Calling#ReAct#MetaGPT#Super-agent#AGENTS.md