Back to home

Compare

Comparing: AI Agents Keep Failing in Enterprise — Not Model IQ, But the Company Itself & AI Agent 在企业落地频频翻车,问题不在模型智力 — 而在公司本身

AEN
Context EngineeringAI AgentOpenAI·

AI Agents Keep Failing in Enterprise — Not Model IQ, But the Company Itself

What This Is

Over the past year, most stories about enterprises deploying AI Agents have shared the same ending: the model runs fine in testing, then derails the moment it hits real business scenarios. The reason usually isn't that the model lacks intelligence — it's that it can't "see" what it needs to see when making decisions.

This pattern has recently been captured by a new term: Context Engineering. It refers to all the information an AI can actually "see" on each call — including conversation history, behavioral rules written by developers, and descriptions of available tools.

Why does it matter? It punctures an illusion. OpenAI researcher Weng Jiayi (翁家翌) is blunt: "Just like people, what matters most for a model is context." Picture an Agent as a genius engineer parachuted into your team — extraordinarily capable, but knowing nothing about your product architecture, business rules, or past decisions. No matter how smart, they can't deliver. That is exactly today's AI Agent predicament.

More importantly, this judgment reframes the bottleneck of AI deployment from a "technology problem" into an "organizational problem": if a team's critical knowledge is tacit and scattered in veteran employees' heads, even the best Agent has nothing to work with.

Industry View

Optimists see this as precisely the trigger for enterprise upgrades. The Linux kernel project has been maintained for over thirty years, with efficient global developer collaboration, precisely because of a highly transparent, documentation-driven culture — these kinds of teams are naturally AI-friendly. In other words, building an AI-native team starts with a documentation movement.

But we must flag the other side. Over-emphasizing documentation may not pay off for small and medium companies: writing and maintaining documentation is itself high-cost, and AI tools iterate extremely fast — today's standards may be obsolete in six months. More subtly, many business judgments rely on tacit understanding and veteran intuition; forcibly documenting them actually strips information out. Treating "context" as a universal cure is another form of techno-optimism.

The judgment worth bookmarking: context is a necessary condition, not a sufficient one.

Impact on Regular People

For enterprise IT: before deploying AI Agents, auditing the company's knowledge assets is more urgent than buying models — otherwise you're paying to hire a brilliant but out-of-touch genius.

For individual careers: deliberately documenting your decision rationale and work logic isn't just good hygiene — it's your personal asset for collaborating with AI in the future.

For the consumer market: the next time you see an AI product touted as "top benchmark scores," ask one more question — in what scenario, and with what information, was it actually tested?

Source: juejin.cn
BZH
上下文工程AI AgentOpenAI·

AI Agent 在企业落地频频翻车,问题不在模型智力 — 而在公司本身

这是什么

过去一年,企业部署 AI Agent 的故事大多有个共同结局:模型在测试里跑得很顺,落到真实业务就走样。原因通常不是模型不够聪明,而是它每次决策时"看不见"该看见的东西。

这件事最近被一个新词概括:上下文工程(Context Engineering)。它指的是 AI 每次调用时实际能"看到"的所有信息——包括对话历史、开发者写好的行为规则、可调用的工具说明。

为什么值得关心?它戳破了一个幻觉。OpenAI 研究员翁家翌的判断很直接:"人和模型一样,最重要的是 Context。" 把 Agent 想象成空降你团队的天才工程师——能力超群,但对你的产品架构、业务规则、历史决策一无所知,再聪明也使不上劲。当下 AI Agent 的困境正是这样。

更重要的,这个判断把 AI 落地的瓶颈从"技术问题"重新定义为"组织问题":如果团队的关键知识是隐性的、散落在老员工脑子里,再好的 Agent 也无从下手。

行业怎么看

乐观派认为,这恰恰是企业升级的契机。Linux 内核项目维护三十多年、全球开发者协作高效,靠的正是高度透明、文档驱动的文化——这种团队天然对 AI 友好。换句话说,建设 AI 原生团队,首先是一场文档化运动。

但我们必须指出另一面。过度强调文档化,对中小公司未必划算:写文档、维护文档本身就是高成本,且 AI 工具迭代极快,今日规范半年后可能过时。更微妙的是,许多业务判断依赖隐性默契和老员工直觉,强行文档化反而会损失信息。把"上下文"当万能解药,是另一种技术乐观主义。

值得记下的判断是:上下文是必要条件,不是充分条件。

对普通人的影响

对企业 IT 来说,部署 AI Agent 之前,先盘点公司知识资产比采购模型更紧迫——否则就是花钱请来一个不接地气的天才。

对个人职场而言,开始有意识地记录自己的决策依据和工作逻辑,这不仅是好习惯,未来也是你与 AI 协作的个人资产。

对消费市场而言,再看到 AI 产品"跑分第一"的宣传,可以多问一句:在什么场景、给了什么信息下测的?

Source: juejin.cn