Armature团队对Claude、Codex和Cursor三个编码智能体进行了大规模对比研究，共进行16893次运行测试，收集了5292个有效会话，涉及51个代码库和18个行业领域1。研究发现这三个智能体在选择第三方服务时采用不同策略，仅在42%的情况下做出相同选择1。
Claude Code倾向于依赖内部知识而较少搜索网络，仅在约30%的会话中进行网络搜索，但在搜索时浏览页面数是Codex的3倍1。相比之下，Codex在94%的会话中使用网络搜索1，而Cursor基于网络做决策的比例为三分之二1。Claude Code自主构建解决方案的能力突出，比例几乎是Codex和Cursor的两倍，达到19%，而后两者均为10%1。
具体的市场选择上，某些服务虽被频繁提及却鲜少被实际采用1。PayPal被提及139次但从未被选中，同类需求中Stripe赢得了其中124个会话1；LangChain被提及194次但仅被选中4次1。在具体领域，Stripe在支付服务上赢率达十分之九1，Neon在数据库领域赢率为66%1，Amazon S3在文件存储领域市场占有率45%1。研究表明编程语言和仓库上下文会显著影响工具选择结果1。
A comprehensive analysis of three major coding agents reveals significant differences in how they choose third-party services when developing software solutions.1 Researchers conducted 16,893 test runs across 51 code repositories spanning 18 industry domains and 10 programming languages, collecting 5,292 valid sessions to examine the decision-making patterns of Claude Code, Codex, and Cursor.1 The three agents selected identical tools in only 42% of cases, demonstrating that their underlying strategies for solving problems diverge considerably.1
The most striking differences emerge in how these agents rely on external information retrieval.1 Codex employed web search in 94% of its sessions, while Claude Code searched the web in approximately 30% of cases but browsed three times as many pages per search compared to Codex.1 Cursor based roughly two-thirds of its decisions on web research.1 Despite these varying reliance levels on external sources, Claude Code independently constructed solutions at nearly twice the rate of Codex and Cursor—19% compared to 10%.1
Specific tool adoption patterns further illustrate how coding agents make pragmatic choices in real-world scenarios.1 Stripe dominated payment service selection with a 9 in 10 win rate, claiming 124 sessions where PayPal was mentioned 139 times but never selected.1 LangChain appeared in 194 discussions yet was chosen only 4 times, while Neon captured 66% of database-related decisions and Amazon S3 held 45% market share in file storage selections.1 These results indicate that mention frequency does not determine adoption, and that programming language, repository context, and problem domain substantially influence tool selection outcomes.1
评论
还没有评论，欢迎留下第一条。