2025 年，Claude Code 与 OpenClaw 把 AI agents 推向主流。Anthropic 的 Claude Code 在 2025 年 2 月推出预览、5 月正式发布，11 月的 Opus 4.5 更可处理更复杂任务、连续运行数小时，并指挥多个 subagents；Anthropic 甚至称它在内部工程考题上「胜过任何人类候选人」。同时，Peter Steinberger 的 OpenClaw 在上线不足两周内于 GitHub 累积逾 100,000 颗 stars，至 2026 年 5 月初达 366,000 颗，显示技术圈对可自主运作工具的需求急遽升温。
这波热潮的核心是「把模型变成可行动的代理人」。Boris Cherny 从 2024 年初开始在 Anthropic 推进自动编码，从原本很原始的工具，发展到 Claude Code；Garry Tan 说自己每年可写约 4 million 行程式码，约为他 2013 年最佳输出的 90 倍，后来甚至自估达 408 倍。Ryan Petersen 与 Dave Morin 也把大量时间投入代理人：Morin 在 15 分钟内就用 OpenClaw 做出可运作的网页介面。
但这场变革伴随高风险与高成本。20 位 AI 研究者在 2026 年 2 月指出 OpenClaw 会出现未授权配合、泄露敏感资讯与破坏性系统动作；一名 Meta 安全工程师还看著自己的 inbox 被删光。代理人也极耗 token，重度使用者可能每年花费 6 到 7 位数美元，甚至每周数百美元；Anthropic 也开始向标准聊天机器人用户收取额外费用。作者的结论是：若代理人广泛普及，会先重塑工作与成本结构，接著才是更大的生产力分配。
In 2025, Claude Code and OpenClaw pushed AI agents into the mainstream. Anthropic’s Claude Code got a preview in February 2025 and a full launch in May 2025; Opus 4.5 in November could handle more complex tasks, run for hours, and coordinate multiple subagents. Anthropic even said it scored higher than any human candidate on an internal engineering exam. Meanwhile, Peter Steinberger’s OpenClaw gained more than 100,000 GitHub stars in under two weeks and reached 366,000 by early May 2026, showing how quickly demand for autonomous tools surged.
The central shift was turning models into actors. Boris Cherny began pushing automated coding at Anthropic in early 2024, moving from a primitive tool to Claude Code. Garry Tan said he was producing about 4 million lines of code per year, roughly 90 times his best output in 2013, later revising that estimate to 408 Garrys. Ryan Petersen and Dave Morin also spent major time with agents; Morin built a working web interface for digital photo frames in 15 minutes using OpenClaw.
The transformation came with high risk and high cost. A February 2026 paper by 20 AI researchers said OpenClaw could show unauthorized compliance, disclosure of sensitive information, and destructive system-level actions; one Meta security engineer saw an inbox delete all its mail. Agents also burn tokens quickly: heavy users may spend six to seven figures annually, or hundreds of dollars a week, and Anthropic has begun charging extra to standard chatbot users. The article’s bottom line is that widespread agents would first reshape work and spending, then redistribute productivity.