被提及194次，被选择4次。
AI编程工具Armature发布了一项大样本研究：16,893次真实会话（其中5292次有效），覆盖1163个prompt变体、75个代码库、10种语言，让Claude Code、Codex与Cursor三个agent在真实任务里自选第三方工具，Gemini 3.7 Flash担任编排与评审。它回答了一个此前从没被数据回答过的问题：AI agent到底怎么选工具？
方法：让agent自己投票
实验设计相当克制：不预设正确答案，只记录agent在搭建支付、数据库、文件存储、邮件等真实集成任务中的搜索、试错与最终选择。先导发现是行为差异——Codex有94%的会话用web搜索，其中九成会加site:操作符定向；Cursor约三分之二用web；Claude Code只有30%用web、但浏览的页面数是别人的3倍，遇到sandbox这类新领域时比例升至80%。三个agent合起来，只在42%的任务类别里选中了同一个工具。
声量与得票，是两张榜
最扎眼的是"提及量"与"得票率"的剪刀差：LangChain在对话中被提及194次，最终只有4次被选；Paypal被提及139次、0次被选——同场景下Stripe赢下124次，支付类胜率90%；Netlify提及152次、6次被选；Supabase被提及242次，却在数据库类被Neon压制（胜率66%）。邮件类里Resend拿35.6%，Postmark 27.4%。还有一个例外值得玩味：Claude Code有19%的任务选择自建方案，Codex与Cursor只有10%—— Sometimes自己动手，胜过任何第三方。
agent的选型逻辑：运行时验证
数据指向一个朴素结论：星标与提及量影响的是训练语料，而agent做决定时更依赖运行时验证——文档能不能读懂、示例能不能直接跑、试错成本高不高。声量大的项目未必赢，文档工程好的项目开始赢。对开发者生态，这是一次选型权重的重估：为agent写作的文档，正在变成新的分发渠道。对那些"被提及却不被选"的项目，问题不在营销，而在于agent上手的那十分钟里，你给它的体验是什么。
agent不投票给声量，投票给它跑得通的东西。
明天见。
Mentioned 194 times. Chosen 4.
AI coding toolmaker Armature published a large-sample study: 16,893 real sessions (5,292 valid), spanning 1,163 prompt variants, 75 codebases and 10 languages, letting Claude Code, Codex and Cursor choose third-party tools on real tasks, with Gemini 3.7 Flash orchestrating and judging. It answers a question no dataset had answered before: how do AI agents actually pick tools?
Method: let the agents vote
The design is restrained: no correct answers are presumed; the study merely logs how agents search, trial and finally choose while building real integrations — payments, databases, file storage, email. A preliminary finding is behavioral divergence: Codex used web search in 94% of sessions, with site: operators in nine out of ten; Cursor used web about two-thirds of the time; Claude Code used web only 30% of the time but browsed 3x as many pages, rising to 80% on novel domains like sandbox. Across all three, agents agreed on the same tool in only 42% of task categories.
Buzz and votes are different leaderboards
The starkest gap sits between mentions and wins: LangChain was mentioned 194 times in conversation, finally chosen 4; PayPal mentioned 139 times, chosen 0 — in the same scenarios Stripe won 124 times, a 90% payments win rate; Netlify was mentioned 152 times, chosen 6; Supabase, mentioned 242 times, got squeezed in databases by Neon (66% win rate). In email, Resend took 35.6% against Postmark's 27.4%. One exception is worth savoring: Claude Code built its own solution in 19% of tasks, versus 10% for Codex and Cursor — sometimes doing it yourself beats any third party.
The agents' logic: runtime verification
The data points to a plain conclusion: stars and mentions shape the training corpus, but agents decide on runtime verification — can the docs be parsed, do the examples actually run, how expensive is trial and error. Loud projects don't necessarily win; well-documented projects start to. For the developer ecosystem, this re-weights tool selection: docs written for agents are becoming a new distribution channel. For the mentioned-but-not-chosen, the problem isn't marketing — it's the ten-minute first-run experience you hand the agent.
Agents don't vote for buzz. They vote for what they can make run.
See you tomorrow.
训练语料记住谁被谈论，运行时验证决定谁被安装——agent把这两张榜拆开了。
—— Dawn Vision编辑部
Training corpora remember who gets talked about; runtime verification decides who gets installed. Agents just split the two leaderboards.
— The Dawn Vision Editorial Desk
Armature 16,893 sessions实测（5292有效）、1163 prompt变体、75代码库、10语言 · 3个agent：Claude Code/Codex/Cursor，Gemini 3.7 Flash任orchestrator+judge · 搜索行为：Codex 94%用web（9/10用site:）；Cursor约2/3用web；Claude Code 30%用web但浏览页面3倍（sandbox新领域升至80%） · 一致性：三agent仅42%任务类别选同一工具 · 剪刀差：LangChain提及194次仅4次被选；Paypal提及139次0次被选（Stripe赢124次，支付胜率90%）；Netlify 152次仅6次；Supabase 242次被Neon压制（66%） · 邮件：Resend 35.6% vs Postmark 27.4% · 自建：Claude Code 19% vs Codex/Cursor 10%
Armature's 16,893-session study (5,292 valid), 1,163 prompt variants, 75 codebases, 10 languages · Three agents: Claude Code / Codex / Cursor, with Gemini 3.7 Flash as orchestrator + judge · Search behavior: Codex 94% web (site: in 9/10); Cursor ~2/3 web; Claude Code 30% web but 3x pages (up to 80% on novel domains like sandbox) · Consistency: same tool chosen in only 42% of task categories · Buzz-vs-votes gap: LangChain 194 mentions / 4 chosen; PayPal 139 / 0 (Stripe won 124, 90% payments win rate); Netlify 152 / 6; Supabase 242 mentions, squeezed by Neon (66%) · Email: Resend 35.6% vs Postmark 27.4% · Self-build: Claude Code 19% vs Codex/Cursor 10%
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理当日采集信号并核对一手研究后生成，经编辑部人工审核。素材来源：Armature研究原文（全部数据与方法论），Hacker News社区讨论（160分热度，未作数据来源）。全部数字来自一手研究，方法与样本量已在正文标注；Hacker News仅作传播热度背景。
This article was generated by the Dawn Vision cognitive engine processing collected daily signals against the primary research, followed by human editorial review. Source: Armature's original study (all figures and methodology); Hacker News discussion (160-point heat) served only as distribution context. All numbers come from the primary research, with method and sample size labeled in the text.
