Anthropic 的 Claude Code 负责人 Boris Cherny 讨论了 AI 代理的演变、对软件工程的影响以及人机交互的未来。
Boris Cherny, head of Claude Code at Anthropic, discusses the evolution of AI agents, their impact on software engineering, and the future of human-computer interaction.
要点 · TL;DR
Claude Code 的成功源于模型改进，而非仅仅是工具本身。 Claude Code's success stems from model improvements, not just the harness.
Anthropic 将 Claude Code 视为安全研究工具，将对齐融入产品开发。 Anthropic treats Claude Code as a safety research tool, integrating alignment into product development.
未来的软件工程将由人类指导 AI 代理，编程语言的重要性将降低。 Future software engineering will involve humans directing AI agents, with programming languages becoming less relevant.
核心观点 · Key points
Claude Code 的增长主要由模型改进驱动，而不仅仅是工具本身。 Claude Code's growth is driven by model improvements, not just the harness.
对齐和安全研究是 Anthropic 产品开发的核心。 Alignment and safety research are core to Anthropic's product development.
软件工程的未来涉及人类指导 AI 智能体。 The future of software engineering involves humans directing AI agents.
企业逐步采用 AI 编码工具，沿着采用阶梯向上移动。 Enterprises adopt AI coding tools gradually, moving up an adoption ladder.
隐私和安全对于企业对 AI 的信任至关重要。 Privacy and security are paramount for enterprise trust in AI.
反共识 · Contrarian takes
Claude Code 是安全研究工具，而不仅仅是商业产品。 Claude Code is a safety research tool, not just a commercial product.
命令行界面是权宜之计，而非最终形态。 The command line interface was a stopgap, not the final form.
对于 AI 生成的代码，编程语言可能变得基本无关紧要。 Programming languages may become largely irrelevant for AI-generated code.
商业中最大的护城河不是转换成本，而是规模效应和网络效应。 The biggest moats in business are not switching costs but scale and network effects.
在某些编码任务中，AI 模型可能比人类更安全。 AI models can be safer than humans in some coding tasks.
本期章节 · Chapters（共 34）
早期原型与计算机使用Early Prototypes and Computer Use
主持人介绍与个人轶事Host Introduction and Personal Anecdote
Claude Code的起源与目的Origin and Purpose of Claude Code
在Anthropic构建编码产品Building a coding product at Anthropic
2026年编码大爆发The 2026 coding explosion
工具的商业目标Business aims of the harness
权限提示与安全演进Permission Prompts and Safety Evolution
规避限制与对齐Circumventing Constraints and Alignment
模型犯错与自我纠正Why Models Make Mistakes and Self-Correct
创意过程与代码质量Creative process and code quality
新范式下工程师角色Role of engineers in the new paradigm
与AI代理协作Collaborating with AI agents
Claude Code界面介绍Introduction to Claude Code interfaces
不寻常的AI工具体验Unusual AI harness experiences
模型发布与商业焦虑Model rollout and business anxiety
算力稀缺与访问担忧Compute scarcity and access concerns
个人与企业使用Claude CodeIndividual vs enterprise usage of Claude Code
采用与安全顾虑Adoption and Security Concerns
模型崩溃与扩展Model Collapse and Scaling
对软件行业的影响Impact on Software Industry
组合模式与前沿模型担忧Combining modes and frontier model concerns
COBOL迁移与语言相关性COBOL migration and language relevance