Claude Code 的起源：从意外到智能代理

The Origin of Claude Code: From Accident to Agent

鲍里斯·切尔尼 Boris Cherny · Acquired Unplugged （WorkOS 呈现） · 2026-06-02 · 约 30 分钟 · 原视频 ↗

本期速览 · Overview

Claude Code 源于 Anthropic 的一个原型团队，其初衷是通过真实的编程交互来研究 AI 安全。

Claude Code originated from a prototyping team at Anthropic, driven by the need to study AI safety through real-world coding interactions.

Claude Code 是意外在 8-9 天内建成的，如今在 Anthropic 带来了巨大的生产力提升。
Claude Code was built accidentally in 8-9 days and now drives massive productivity gains at Anthropic.
模型改进而非智能体框架是 Claude Code 性能提升的主因。
Model improvements, not agent scaffolding, were the main driver of Claude Code's performance.
未来的工程将是编写循环来提示模型，而非直接写代码。
Future engineering will involve writing loops that prompt models, not writing code directly.

Claude Code 起源于 Anthropic 的一个原型团队，作为编码智能体构建，以推动模型前沿。
Claude Code originated from a prototyping team at Anthropic, built as a coding agent to push model frontiers.
编码是 AI 安全研究的干净领域，因为它有明确的通过-失败结果和受限的解决方案。
Coding is a clean domain for AI safety research because it has clear pass-fail outcomes and constrained solutions.
模型改进，尤其是 Sonnet 4 和 Opus 4.5，是 Claude Code 性能提升的主要驱动力。
Model improvements, especially Sonnet 4 and Opus 4.5, were the primary driver of Claude Code's performance gains.
在 Anthropic，Claude Code 的使用使每位工程师的代码量增长数百个百分点，并将入职时间缩短至 2 天。
At Anthropic, Claude Code usage has led to a many-hundred-percent increase in code per engineer and reduced ramp-up time to 2 days.
工程的未来涉及编写循环来提示模型，而不是直接编写代码或提示。
The future of engineering involves writing loops that prompt models, not directly writing code or prompts.

Claude Code 是偶然产物，在 8-9 天内建成，最初只编写 10-20% 的代码。
Claude Code was an accident, built in 8-9 days, and initially only wrote 10-20% of code.
Anthropic 使用扁平化头衔如“技术员工”以避免对资历的遵从并鼓励反驳。
Anthropic uses flat titles like 'Member of Technical Staff' to avoid deference to seniority and encourage pushback.
目前被视为优势的产品品味将随着模型生成好想法的能力增强而消失。
Product taste, currently considered alpha, will erode as models get better at generating good ideas.
人类的最终角色是教授模型价值观，类似于教导孩子如何成为好人。
The ultimate human role is teaching models values, similar to teaching children how to be good people.
用更少的人和更多的代币来资助项目可以带来复合效率提升。
Underfunding projects with fewer humans but more tokens can lead to compounding efficiency gains.