你愿意把多少控制权交给一个大语言模型,让它掌管你的数字生活?
要从模型中获取最大价值,就意味着要把“钥匙”交给它。对于控制欲强的人或对 AI 持犹豫态度的人来说,这似乎有些过头。但对于 OpenAI 桌面应用的首席工程师 Andrew Ambrosino 而言,这是测试未来的唯一方式,正因如此,该应用如今已经可以访问并控制他的收件箱、Slack 账号、手机,以及 Notion、Figma 等应用。
“如果我要它写一份文档,它有没有可能从某个私人私信中提取相关内容,却不知道某些信息不该被分享?有可能,”Ambrosino 告诉 TechCrunch。“为了工作我会这么做。如果有必要,我愿意在个人层面偶尔承担一些风险。而到目前为止,我还没遇到过这种情况。”
Ambrosino 参与开发的是 OpenAI 最大的押注——ChatGPT Work。该产品于上个月发布,可在该公司最低订阅档位使用,每月 20 美元。该产品旨在让白领工作者能够调度 AI 智能体——将大语言模型接入会计师、投资者、医生以及其他所有日常工作以电脑为核心的人群所使用的数字工作流。
OpenAI 的营销文案简明扼要地概括了这一目标:一个“‘人工’智能不再止步于回答问题,而是帮助每个人把最宏大的想法变成现实”的世界。
对于软件开发人员来说,这种转变已经发生,但向其他部门的扩散仍然缓慢。ChatGPT Work 是该公司 Codex 编程工具的修改版本。它的目的是让非工程师也能获得软件工程师已经从智能体中获得的同等功能:一个不仅能回答问题,还能自主完成多步骤项目的 AI 工具。
“在这个新形态下,ChatGPT 实际上可以完全自主地为你完成整个、非常复杂的任务,而且方式令人愉悦且安全,”OpenAI 核心产品工作(包括 Work)负责人 Thibault Sottiaux 告诉 TechCrunch。“这正是 OpenAI 的使命——让所有人都能跟上。”
从商业角度看,这一点至关重要。智能体工作的时间越长,消耗的模型 token 就越多,这使得它们在每位用户身上为 OpenAI 带来的收益更高。触达新的职业领域至关重要——这不仅对 OpenAI 如此,对整个行业也是如此。如果说编程已被证明是 AI 实验室的盈利领域,那么它仍然只是 AI 工具需要支撑的专业工作中极小的一部分,这些公司才能证明其在训练和算力上的巨额投入是合理的。虽然实验室一直专注于软件工程师,但像 Harvey(面向法律)和 Clay(面向销售)这样的垂直领域竞争对手,一直在以模型无关的方式争夺这些客户,这意味着他们会接入当时效果最好的任何 AI 模型。
行业分析师认为,这是 OpenAI 及其竞争对手面临的主要挑战之一。“如果实验室无法迅速掌握在市场上规模化 AI 所需的关键互补性资产,价值将流向其他地方,”Christian Catalini 在 a16z 的“It’s time to build”博客上写道。
让 AI 应用为并非软件工程师的人群所用,需要更多的引导和支持。OpenAI 的非工程人员,比如通讯和财务团队,开始使用 Codex 时,“它对他们并不友好——会问他们代码相关的问题,并显示‘哦,你这个东西的 diff 是空的’,”Ambrosino 说,他指的是为软件变更设计的技术读数。“所以,从二月到现在,我们开始让它变得更加通用。”

OpenAI 支持的一项研究发现,6 月份,98% 的 OpenAI 员工在使用 Codex,但只有 17% 的组织订阅者和不到 1% 的个人订阅者在使用这款智能体编码工具。公司内部近乎全员采用与外部几乎无人问津之间的这种差距,既是这家公司面临的挑战,也是机遇。“我们为用户创造的价值和效用越多,他们就越愿意为其中一部分效用付费,我们对 ChatGPT 的看法也一直如此,”Sottiaux 说。“你坐在那里就会想,‘我当然愿意为此每月付 20 美元’,因为你获得的价值远远超过这个数。”
如何让 AI 变得直观易用
要理解这种脱节,有必要先了解 OpenAI 的工程师们在构建什么。每个大语言模型都需要工程师所说的“套件”——包裹在模型周围的软件,它决定模型能看到哪些信息、能使用哪些工具,以及如何向你呈现答案。
如果你想让那个模型去做事——成为智能体——套件就会为它提供工具,以及如何在长期任务中使用这些工具的指令。对于开发者来说,一个能让大语言模型编写代码的命令行界面(CLI)就足以改变软件的构建和部署方式。但大多数人并不使用 CLI;Windows 取代 DOS 是有原因的。
一个超越软件工程的智能体产品“将不得不应对你生活中杂乱无章的世界、你的工具,以及那些建于 1995 年、从未更新过的网站,”Ambrosino 告诉 TechCrunch,并解释说,他的团队正在构建的这些体验对于扩大实用 AI 的普及范围至关重要。
以 Claude Code 和 Codex 这类应用为例:它们通过抽象掉所有实际的软件编写工作,让用户只需告诉模型他们在程序中想要什么,从而释放了“氛围编程”的潜力。现在,OpenAI 希望让 OpenClaw 这类工具中的功能——程序员用它们来让大语言模型干活——变得像输入提示词一样简单。
安布罗西诺表示:“如果模型面前没有这些产品,专家们知道如何获得同样的结果,但你不可能让十亿人都用上这个东西。”这种在高级用户需求与主流采用要求之间的取舍,体现在OpenAI内部的争论中——一些员工认为,如果用户可以直接向模型提问,按钮就没有必要存在。
“我们对此持反对意见——因为现在还为时过早,”安布罗西诺说。“在这个阶段,可发现性很重要,而到某个时点我们就不需要按钮了。”
Work有更多用于选择项目和插件的按钮,但其目标是与OpenAI其他产品相同的魔法盒式界面。他将其比作拟物设计——这种日渐式微的做法是让数字工具看起来像它们所取代的实体物品,比如让计算器应用看起来像袖珍计算器。“那些东西不只是尴尬的设计。它们实际上帮助人们进入这个领域,并完成过渡,”安布罗西诺说。
OpenAI没有透露有多少人使用Work而非Codex,但这款合并后的应用仅有2000万用户,而该公司称ChatGPT在线用户超过10亿。
给ChatGPT一个技能许可
目前,OpenAI将这款工具定位为最适合常规、数据密集型的协调任务。例如,其员工正在用它设置每周指标报告,并把电子表格变成规划工具。
我与一些风险投资人交流过,他们用智能体将相关的沟通和分析汇编成投资备忘录,还有运营团队用它搭建定制化仪表盘和数据可视化。Sam Altman用它来规划假期。一位OpenAI工程师描述说,他让程序查看一段关于工程问题的Slack对话并“制作一些图表”,然后收到了一系列富有洞察力的图表。
“对于这些公司的普通员工,包括我自己在内,信息量都太大了,”OpenAI 产品工程团队负责人 Akshay Nathan 表示。“我们实际上受限于自身解析所有可得信息并据此采取行动的能力。这些信息存在于所有这些系统记录工具中(比如 Salesforce……ChatGPT 的价值在于你本来就能访问这些信息,但现在你真正能够利用它们了。”
因此,这或许正是 AI 布道者们梦寐以求的数字个人助理。与 Claude Cowork 或 Perplexity AI 浏览智能体类似,ChatGPT Work 将智能体连接到你的现有工作空间——电子邮件、网页浏览器、一系列 SaaS 平台——并让这些上下文为你所用。
当系统正常运转时,效果令人印象深刻:我让 ChatGPT Work 从我的邮箱里找出我儿子那格式奇怪的幼儿园日历,并导入到我的 Google Calendar 中,它做到了,省去了我大量重复的数据录入工作。希望这样我就不会再忘记学校的聚餐活动,或者没能安排好假期的托儿事宜。
我不放心让 OpenAI 访问我的收件箱、采访来源或稿件草稿(AI 反对者们不必担心),也不会让它访问我的银行账户,但我相信如果我有这份信任,它会更有用。我让它对我报道的上市公司做财务分析,它为我生成了一个自动更新的指标仪表盘;它还帮我建立了一个可查询的太空发射数据库,这项任务以前我得靠写 Python 脚本才能完成。它还会每周给我发一封关于学术预印本平台上新发表 AI 研究的邮件。我会继续尝试使用它。
虽然向模型提出请求很直观,但为它提供采取行动所需的权限却没那么简单。为智能体设置访问权限(比如访问云盘)既令人困惑又绕来绕去——我多次尝试只给它“只读”权限,却一直收到错误提示。模型本身帮不上什么忙,但最终在移动应用上,弹出一个对话框告诉我,只有完全访问权限才能让它正常工作。
许多重要设置只能在网页应用中使用,所以我经常发现自己同时在这两个平台上工作。有时 ChatGPT Work 的局限性令人费解——将其关联到你的 Google 日历后,它可以创建事件,但不能创建新日历。而且,除非把努力程度调高,否则别指望它能做好任何事,不然你就像在跟一个你遇到过的最差劲的实习生共事。
这是 AI 早期采用者给出的常见建议,他们担心受挫的新手会放弃。OpenAI 的 harness 工程负责人 Joe Gershenson 承认,努力程度设置对新用户来说还不够直观——“我们可以在帮助他们获得合适的推理水平方面做得更好……”他说,并补充道,“敬请期待。”
OpenAI 在打入普通白领工作领域时还面临另一个重要挑战:大多数工作流程不像代码那样可衡量或可评估。软件要么能用要么不能用,即便这种区分也简化了代码优劣的细微差别。而一份好的演示文稿、商业策略或销售说辞则没那么容易评估或追溯。
“像这样的产品面临的独特挑战之一就是它真的什么都能做,”Ambrosino 说。当我问到团队围绕哪些具体问题来设计、针对哪些工作流程时,与我交谈的工程师们婉拒了回答,称这是应该问 OpenAI 研究团队的问题。
OpenAI 后来给出了答案,告诉 TechCrunch 它使用自己的基准测试 GDPVal,该基准取自 44 个职业和数百项知识工作测试,并以用户反馈作为补充。一个不那么官方的答案是,这些设计思路来自 OpenAI 员工自身。Ambrosino 说:“我们必须不断分辨……我们正在做的工作流程是其他人都会做的,还是我们比较另类?”
该应用的早期采用者将通过他们的实际使用留下有价值的痕迹——编程工具的成功很大程度上就建立在类似的数据收集之上——前提是他们不选择退出将数据用于训练。(我就选择了退出)。
推动 OpenAI 产品设计的竞争
尽管外界对模型界面的关注度很高,OpenAI 的工程师们却不愿回答一个相当简单的问题:Codex 和 ChatGPT Work 与 Claude CoWork,或其他面向大众用户群体的竞品智能体框架相比,究竟有何不同?“这个答案可能会让你非常失望,我很抱歉,但诚实的回答是,我真的不看他们构建的那些框架,”Gershenson 在一个典型的回答中说道。“《广告狂人》里那句‘我根本就没在想你’的梗,用在这里很贴切。”
坦白说,我不太相信他们的话,单凭产品用户界面之间的极度相似性、任何企业都需要竞争情报这一点,以及我启动 ChatGPT Work 时它让我做的第一件事就是把我的 Claude Cowork 数据迁移过来,就足以让我怀疑。
如果 Claude Code 在 OpenAI 办公室里是个敏感话题,那也是可以理解的。他们的企业竞争对手定义了 AI 编程市场,并掀起了一场软件工程师工作方式的革命。更让人恼火的是,OpenAI 最初有这个想法,但没能正确地驾驭它。
当 OpenAI 最初将 Codex 开发为网页应用时,工程师们有点操之过急——或者,用 Ambrosino 的话说,“有点过于沉迷 AGI 了。”简而言之,他们押注于模型足够聪明,能在用户输入极少的情况下完全独立处理任务。
紧随其后构建的 Claude Code,则围绕与用户来回对话的模式展开。如果你给它一个问题,它会审视各种可能性,并给你三到四个可选的推进方案。一旦你做出选择,它会再推进一点,然后再次回来确认,持续提供更新,从而给模型和框架留下更少的出错空间。
Anthropic 的做法被证明更有效,尽管它对用户提出了更多要求。“(我们的)产品当时比模型和框架的发展水平稍微超前了一点,”Ambrosino 现在这样说道。
OpenAI 最终也紧随其后,为用户增加了更多与模型互动的机会。这便演变成了我们今天所熟知的 Codex,并推出了桌面端和移动端应用。以下载量作为衡量用户对程序兴趣的指标,Claude Code 在今年四月之前一直更受欢迎,但如今 Codex 已略微领先;对企业使用情况的调查也表明 OpenAI 正在迎头赶上。
这种领先部分源于找准了产品市场契合点,部分则源于用户对 Anthropic 模型安全限制的抱怨以及算力短缺问题。OpenAI 通过 ChatGPT Work 继续朝着更以人为中心的工具方向迈进,但受访的工程师坚称,关键差异在于 OpenAI 最新强大且高性价比模型的实力。
“令人沮丧的答案是,很多时候关键就在于模型本身,而我们在打造这款应用时一直努力做好的一件事,就是充分利用模型的能力,”Ambrosino 说道。
那么,究竟什么才算一个好的工具框架呢?
这个问题的答案要回归到 AI 研究者所领悟的“苦涩的教训”:一个更好的通用模型比特定的领域经验更为重要。对于真正的信奉者而言,工具框架只是临时的拐杖,而非护城河。
“短期内,你可以通过添加一大堆额外的东西——各种条件判断和工具——来获得不错的结果,但说真的,下一代模型几个月后就会发布,让这一切变得过时,”Gershenson 告诉 TechCrunch。他的团队专注于用最简单的方式,将模型所需的工具和上下文暴露给它——仅此而已。
“优秀工具框架工程的目标是……更精确地确定模型真正需要哪些信息来解决你的问题,因为只要你放手让模型去做,它们在这方面的能力会越来越强,”Gershenson 说道。
不过,一个悬而未决的问题是,大多数人或模型是否已经为此做好了准备。沃顿商学院研究职场 AI 工具的教授 Ethan Mollick 仍然认为 Claude 更易于使用,他写道:“ChatGPT 倾向于施展‘魔法’并直接为你完成,而 Claude 则会进行对比并展示出来,反复征求你的意见和反馈,并进行 A/B 测试。”
索蒂奥(Sottiaux),或许整个 OpenAI 也持同样看法,认为应用的对话式特性比学习如何使用一款应用更好。“我们确实看到世界似乎已经准备好了,”他在谈到这款应用时表示。“这就是为什么我们获得了惊人的采用率。”
尽管如此,针对特定模型定制的“驾驭框架”(harness)是否就是最大化模型性能的正确押注,目前仍不明朗。Composio 和 Databricks 等公司进行的对比测试显示,不同的框架与模型组合在编程基准测试上表现各异。Databricks 发现,由软件公司 Earendi 发布的开源框架 Pi,在使用同一个 GPT 5.5 模型时,性能优于 Codex。Pi 已被用于构建 OpenClaw 和 CloudflareOS 等项目。
Pi 的创造者马里奥·泽赫纳(Mario Zechner)表示,他这款刻意走极简路线的框架证明了,一种“AGI 信仰”式的路径是可行的,至少对软件工程师和编程任务而言如此。他说,它在显性功能上的不足,由它自我修改、构建自身界面的能力来弥补。他对 OpenAI 工程师在将用户群拓展到工程师之外时所面临的挑战感同身受。
“一切都被塑造成编程智能体的形态……原因在于他们只有编程智能体任务的训练数据,”他告诉 TechCrunch。“假设我身处管理层,我今天做了一个决定,结果几个月后才显现。你无法在用户与智能体之间简单的来回交互记录中捕捉到这一点,所以所有这些类型的任务,以及任何你未数字化的东西,模型都无法学习。”
与其他开源提供商一样,他认为大型实验室推广自家框架的努力是一种锁定用户的方式;“他们需要拥有整个技术栈;否则,他们就只能沦为模型提供商,然后不得不与中国模型竞争。”
他以及 TechCrunch 采访到的其他工程师都认为,前沿实验室框架中关于 token 消耗和智能体行为的洞察过于有限。从某种意义上说,这对非技术工作者意义较小,但正如编程工具的情况一样,一旦采用规模达到 OpenAI 所希望的量级,最终将迫使人们就成本问题展开更艰难的对话。
举个例子,我在一个每月20美元的订阅套餐里随意使用,四天内就用掉了超过8000万token,根据模型自己的分析(应用里没有仪表盘),这花费了65美元。也就是说,仅四天的随意使用,补贴就超过了订阅价格的3倍。
“我们每天都在努力推动效率的边界,”Sottiaux说,他提到OpenAI的Luna模型最近降价80%。“如果你六个月后醒来,你应该能用更少的开销完成所有同样的事情。”
另一个相关的问题是,这些应用是否会通过数据留存,或者配置所有插件及其权限访问的繁琐过程,对客户产生锁定效应。
我七月份参观了OpenAI那间铺着木饰板、摆满绿植的总部,那里的氛围平静中带着一丝紧张;这些人手头有太多事要做。和我交谈的工程师们在谈话时不停地盯着笔记本电脑,从一个会议室匆匆赶到另一个会议室。
产品工程团队负责人Nathan表示,重点仍然是“那个魔法盒子的承诺,但我仍然觉得复杂度太高了……我非常乐观地认为我们可以解决这个问题,用模型本身,以一种真正AI原生的方式。”
How much control are you willing to give an LLM over your digital life?
Getting the most value from a model means giving it the keys. For a control freak or the AI-hesitant, it seems like a lot. For Andrew Ambrosino, the lead engineer for OpenAI’s desktop app, it’s the only way to test the future, which is why that app now has access to, and control over, his inbox, his Slack account, his phone, apps like Notion and Figma, and more.
“If I’m asking it to write a document, is there a possibility that it’s going to pull from a private DM on that subject and not know that it’s not supposed to share some info? Yes,” Ambrosino told TechCrunch. “I’ll do it for the job. I will take the personal hit here and there if I have to. And I haven’t had to.”
Ambrosino works on OpenAI’s biggest bet, ChatGPT Work, which was released last month and is available on the company’s lowest subscription tier, for $20 a month. The product is intended to allow white-collar workers to field AI agents — hooking LLMs up to the digital workflows used by accountants, investors, doctors, and everyone else whose day-to-day is dominated by their computer.
OpenAI’s marketing copy puts the goal succinctly: A world where “where [artificial] intelligence goes beyond answering questions to helping everyone turn their biggest ideas into reality.”
For software developers, that shift is already happening, but it’s been slow to spread to other departments. ChatGPT Work is a modified version of the company’s Codex coding tool. It’s meant to give non-engineers a version of the same functionality that software engineers already get from agents: an AI tool that doesn’t just answer questions, but completes multistep projects on its own.
“In this new factor, ChatGPT can actually do entire, very complicated tasks for you all autonomously in a way that is delightful and safe,” Thibault Sottiaux, who leads OpenAI’s core product work, including Work, told TechCrunch. “It’s the very mission of OpenAI — to bring everyone along.”
Commercially, that matters a lot. Agents that work for longer stretches burn through more tokens, which makes them more lucrative for OpenAI on a per-user basis. Reaching new professions is crucial — not just for OpenAI, but for the industry at large. If coding has proven lucrative territory for AI labs, it’s still a tiny subset of the professional work AI tools need to enable if these companies are to justify their massive investment in training and computation. While labs have been focused on software engineers, vertical-specific competitors like Harvey (for law) and Clay (for sales) have been chasing those customers with a model-agnostic approach, meaning they’ll plug in whichever AI works best at the time.
Industry analysts see this as one of the major challenges facing OpenAI and its competitors. “If the labs cannot rapidly get ahold of the key complementary assets needed to scale AI in the market, value will accrue elsewhere,” Christian Catalini wrote on a16z’s “It’s time to build” blog.
Making the AI apps work for people who aren’t software engineers requires more hand-holding. OpenAI’s non-engineering workforce, like the communications and finance teams, started using Codex “at a time that it was actively hostile to them—asking them about code and showing them, ‘oh, you have an empty diff for this thing,’” Ambrosino said, referring to a technical readout meant for software changes. “So, we started to make it more general purpose between February and now.”

An OpenAI-backed study found that in June, 98% of OpenAI employees were using Codex, but just 17% of organizational subscribers and less than 1% of individual subscribers were using the agentic coding tool. That difference between near total adoption inside the company and negligible adoption outside it is the challenge and opportunity for the company.
“The more value and the more utility that we generate for users, the more they will be willing to also pay for some part of that utility, and that’s how we’ve always seen ChatGPT as well,” Sottiaux said. “You sit there and you’re like, ’of course I want to pay $20 bucks a month for this,’ because the value that you get is so much more.”
How to make AI intuitive
To understand that disconnect, it helps to understand what OpenAI’s engineers are building. Every LLM requires what engineers call a “harness” — the software wrapped around a model that decides what information it sees, which tools it can use, and how it presents its answers back to you.
If you want that model to do stuff — to become an agent — the harness gives it tools and instructions for using them on long-term tasks. For developers, a command-line interface (CLI) that enabled LLMs to code was enough to change the way software was built and deployed. But most people aren’t using CLIs; there’s a reason Windows replaced DOS.
An agentic product that goes beyond software engineering is “going to be something that plays with the messy world of your life and your tools and websites that were built in 1995 and never updated,” Ambrosino told TechCrunch, explaining that the experiences his team is building are vital to expanding access to useful AI.
Consider apps like Claude Code and Codex: They unleashed “vibe coding” by abstracting away all the actual software writing, and letting users just tell the model what they want in a program. Now, OpenAI wants to make functionality found in tools like OpenClaw, which coders use to put LLMs to work, as easy as prompting.
“Without these products in front of the model, experts would know how to get the same results, but you wouldn’t get to a billion people using the thing,” Ambrosino said. That trade-off between what power users need and what mainstream adoption requires plays out in internal debates at OpenAI, where some employees argue that a button is unnecessary if users can just ask the model directly.
“We push back on [that] — because it’s very early,” Ambrosino said. “Discoverability matters in this phase, and at some point we won’t have the button.”
Work has a few more buttons for selecting projects and plug-ins, but it aims for the same magic box interface as other OpenAI products. He compares it to skeuomorphism, the fading practice of making digital tools look like the physical objects they replaced, like a calculator app made to look like a pocket calculator. “That stuff wasn’t just cringe design. That actually helped get people into this [and] make the transition,” Ambrosino said.
OpenAI wouldn’t say how many people used Work versus Codex, but the joint app is used by just 20 million people, compared to more than a billion users the company says are prompting ChatGPT online.
Giving ChatGPT a license to skill
For now, OpenAI is pitching this tool as best suited for routine, data-intensive coordination tasks. Its employees are setting up weekly metrics reports, for example, and making spreadsheets into planning tools.
I’ve spoken to VCs using agents to assemble relevant communications and analysis about companies into investment memos, and ops teams spinning up bespoke dashboards and data visualizations. Sam Altman is using it to plan his vacations. One OpenAI engineer described asking the program to look at a Slack conversation about an engineering problem and “make some charts,” then receiving back a series of insightful plots.
“There is a deluge of information for the average worker or employee of any of these companies, including myself,” Akshay Nathan, who leads the product engineering team at OpenAI, said. “We’re actually quite limited by our ability to parse everything that’s available to us, and then take action on it. That information lives in all these system records tools [like, Salesforce…the value of ChatGPT is you already have access to this, but now you truly have access to it.”
This, then, could be the digital personal assistant that AI evangelists dream about. As with Claude Cowork or Perplexity AI browsing agent, ChatGPT Work links agents to your existing workspace — email, web browser, a slew of SaaS platforms — and puts that context to work for you.
When the system works, it can be impressive: I asked ChatGPT Work to get my son’s weirdly-formatted preschool school calendar out of my email and put it into my Google Calendar, and it did, saving me a lot of repetitive data entry. Hopefully now I won’t forget the school potluck or fail to arrange vacation childcare.
I didn’t trust OpenAI with access to my inbox, source interviews or story drafts (fear not, AI haters) and wouldn’t let it have access to my bank account, but I believe it would be more useful had I the faith. I tasked it to do financial analysis on publicly-traded companies that I cover, and it delivered an auto-updating dashboard of metrics for me; it made a queryable database of space launches, a task I’d previously had to accomplish by writing python scripts. It also sends me a weekly email about new AI research posted at academic clearinghouses. I’ll keep experimenting with it.
While asking the model for something is intuitive, giving it what it needs to take action isn’t as simple. Setting up the permissions for agents to access, say, a cloud drive was confusing and circular — I tried multiple times to give it just “read” access and received error messages. The model itself wasn’t too helpful, but eventually on the mobile app, a dialog box popped up to tell me that only complete access would make it work.
Many important settings are only available on the web app, so I frequently found myself working in both at the same time. Sometimes ChatGPT Work’s limitations are baffling—link it to your Google calendar and it can create events, but not new calendars. And don’t bother trying to do anything unless the effort level is high, otherwise you’ve got the worst intern you’ve ever worked with.
That’s common advice from AI early adopters, who fear that frustrated newbies will give up. Joe Gershenson, the engineering lead for OpenAI’s harness, admitted that effort settings aren’t intuitive for new users yet—”there are things that we can do better to help them get the right level of reasoning…” he said, adding, “Watch this space.”
OpenAI faces another important challenge breaking into normie white-collar work: Most workflows aren’t as measurable—or evaluable—as code. Software either works or it doesn’t, and even that distinction reduces the nuance about what makes code good or bad. A good presentation, business strategy, or sales pitch isn’t as easy to evaluate or trace.
“One of the unique challenges with a product like this is just that it can really do anything,” Ambrosino said. When I asked what specific problems the team designs around, and what workflows it targets, the engineers I spoke with demurred, saying that was a question for OpenAI’s research team.
OpenAI later provided an answer, telling TechCrunch that it uses its benchmark GDPVal, drawn from 44 occupations and hundreds of knowledge work tests, and supplements that with user feedback.A less official answer is that it comes from OpenAI employees themselves. Said Ambrosino, “We have to always parse out … are we doing the workflow that everybody else will be doing, or are we weird?”
The early adopters of the app itself will create valuable traces with their actual usage—much of the success of coding tools is built on similar data collection—assuming they don’t opt out of making it available for training. (I did).
The rivalry that drove OpenAI’s product design
Despite all the attention on the model interface, OpenAI’s engineers were reluctant to answer a fairly simple question: What sets Codex and ChatGPT Work apart from Claude CoWork, or other competing agentic harnesses intended for a mass user base?
“It’s going to be a really disappointing answer for you, and I’m sorry, but the honest answer is that I really don’t look at the harnesses that they’re building,” Gershenson said in a typical answer. “The Mad Men ‘I don’t think about you at all’ meme comes to mind here.”
Frankly, I don’t believe them, if only based on the extreme similarities between the products’ user interfaces, the need for competitive intelligence at any business, and because the first thing ChatGPT Work asked me to do when I started it up was port over my Claude Cowork data.
It’s understandable if Claude Code is a sensitive topic around the OpenAI offices. Their corporate rival defined the market for AI coding and launched a revolution in how software engineers do their jobs. It’s additionally frustrating because OpenAI had the idea first, but didn’t quite harness it correctly.
When OpenAI first developed Codex as a web app, the engineers got a bit over their skis—or, as Ambrosino puts it, “a bit more AGI-pilled.” In short, they bet on the model being smart enough to handle a task entirely on its own, with minimal user input.
Built shortly afterward, Claude Code was oriented around a back-and-forth conversation with the user. If you gave it a problem, it would survey the possibilities and give you three or four options for proceeding. Once you chose, it would go a little further and then check back again, giving continual updates and leaving less room for the model and harness to make mistakes.
Anthropic’s approach proved more effective, even if it demanded more work from users. “[Our] product was a little ahead of where the model and harness was at the time,” Ambrosino says now.
OpenAI eventually followed suit by adding more opportunities for users to interact with the model. That became the Codex we know today, with desktop and mobile apps. Using download statistics as a proxy for interest in the programs, Claude Code was more in demand until April of this year, but now Codex has taken a slight lead; surveys of enterprise use also suggest OpenAI is catching up.
Part of that lead is getting the product-market fit right, and part comes from complaints about safety restrictions on Anthropic’s models and compute shortages. OpenAI’s steps toward more human-centric harness continue with ChatGPT Work, but the engineers I spoke to insisted the key differentiator is the strength of OpenAI’s latest powerful and cost-effective models.
“The frustrating answer is that a lot of times it is the model, and one thing that we have tried to do really well with this app is fully leverage the model,” Ambrosino said.
What makes a good harness, anyway?
That explanation returns to the “bitter lesson” learned by AI researchers that a better general model is more important than specific domain experience. For true believers, the harness is a temporary crutch, not the moat.
“You could get good results in the short term by adding a whole bunch of extras—if and thens and tools—but like, come on, the next model is going to come out in a couple of months and make that obsolete,” Gershenson told TechCrunch. His team focuses on the simplest ways to expose the model to the tools and context it needs—and no more.
“The goal of good harness engineering is to… be more precise about what information the model really needs to solve your problem, because the models are getting better and better at doing that if you simply let them do their thing,” Gershenson said.
There is an open question, though, if most people or models are ready for that. Ethan Mollick, the Wharton School of Business professor who studies AI tools in the workplace, still sees Claude as more user-friendly, writing that “ChatGPT tends to want to do magic & just do it for you, while Claude does comparisons & shows them, repeatedly asking for input & feedback and doing A & B tests.”
Sottiaux, and perhaps OpenAI at large, disagree, arguing that the conversational nature of the app is better than learning how to use an application. “We definitely see that the world seems to be ready,” he says of the app. ”This is why we’ve had incredible adoption.”
Still, it’s not clear that a model-specific harness is even the right bet for maximizing a model. Comparisons run by companies like Composio and Databricks show that different harness and model combinations deliver different performance on coding benchmarks. Databricks found that Pi, an open-source harness published by the software company Earendi, outperformed Codex while using the same GPT 5.5 model. Pi has been used to build projects like OpenClaw and CloudflareOS.
Pi’s creator, Mario Zechner, says his intentionally minimalist harness is evidence that an AGI-pilled approach can work, at least for software engineers and coding tasks. What it lacks in explicit features, he says, is made up for by its ability to modify itself and build its own interfaces. He sympathizes with the challenge that OpenAI’s engineers face in expanding their user base beyond engineers.
“Everything is coding agent shaped…the reason is that they only have training data for coding agent tasks,” he told TechCrunch. “Say I’m in management, I make a decision today, and the outcome happens months later. You cannot capture that in a simple trace of a user and agent back and forth, so all of these kinds of tasks and anything that you don’t digitize is inaccessible to a model to learn.”
Like other open source providers, he sees the big lab’s effort to push their harnesses as a way to lock-in users; ”They need to own the entire stack; otherwise, they just become a model provider and then need to compete with Chinese models.”
He and other engineers TechCrunch spoke to felt that the insight into token spend and agent behavior in frontier labs’ harnesses is too limited. In a sense, that’s less meaningful to non-technical workers, but as with the coding tools, uptake at the scale OpenAI hopes for will eventually force harder conversations about cost.
For example, messing around on a $20-a-month subscription, I used more than 80 million tokens in four days, which cost $65, according to the model’s analysis (there’s no dashboard in the app). That’s a subsidy of more than 3x the subscription price for four days of casual use alone.
“We are working every day to push the frontier on efficiency,” Sottiaux said, pointing to a recent 80% price cut for users of OpenAI’s Luna model. “If you wake up six months from now, you should be able to do all of the same [tasks] with less spend.”
The other relevant question is whether these apps create a lock-in effect on customers through data retention, or the sheer pain of configuring access to all the plug-ins and their permissions.
Inside OpenAI’s wood-panelled, plant-filled headquarters, which I visited in July, the atmosphere was calm but slightly tense; these are people with a lot to do. The engineers I spoke with were constantly monitoring their laptops as we talked, and rushing from meeting room to meeting room.
Nathan, the head of the product engineering team, said the focus remains on “the promise of the magic box, but I still think there’s too much complexity…I’m very optimistic that we can solve it, with the model and in a truly AI native way.”