加载中...

上周,一家中型初创公司的工程负责人警告说,他的团队在短短两个季度内,编码代理的账单膨胀了六倍——不是因为工作变得更难,而是因为没有人关注。LangChain 博客指出,同样的模式如今也出现在 Uber、微软和 Salesforce 等公司,不受控制的 token 支出正在将前景光明的 AI 试点变成预算噩梦。
根据 LangChain 博客,2026 年初编码代理的使用量激增,团队错误地将更高的 token 消耗等同于进展。几个月内,Uber 在四个月内耗尽了其 2026 年的全部 AI 预算,微软开始取消各部门的 Claude Code 许可证,而 Salesforce 正面临 3 亿美元的 Anthropic 账单。博客解释说,真正的问题是成本可见性碎片化:每个工具(如 Claude Code、Cursor、Copilot Chat、Pi、OpenCode 等)都以自己的格式记录使用情况,这使得回答诸如“我们实际花了多少钱来构建这个功能?”这样简单的问题变得不可能。没有统一的视图,token 最大化将成为一种负担,而不是杠杆。
LangChain 团队为运行多个编码代理的团队构建了解决方案。他们的平台 LangSmith 将所有受支持代理的跟踪数据规范化为通用模型(根会话、轮次、工具调用、元数据)。这使得团队能够按线程 ID、模型、提供商或工具名称进行筛选,并在不切换仪表板的情况下比较不同代理中昂贵或失败的会话。博客强调,可观测性是第一步——你需要知道哪些代理正在运行,它们花了多少钱,以及会话在哪里失败,然后才能优化或管理支出。后续阶段——“引擎”和“LLM 网关”——接入相同的跟踪数据,以强制执行上限并自动修复问题,而无需拆除现有集成。
如果你正在管理一个精简的工程团队,你浪费的每一个 token 都是你无法重新投资于产品或招聘的美元。LangChain 的案例研究表明,即使多一个工具也可能让预算失控,因为原生仪表板只能看到自己的一小部分。通过 LangSmith 整合可观测性,你可以获得所有编码代理活动的单一事实来源,将不透明的支出转化为可操作的数据。这很重要,因为它让你能够:
简而言之,没有跨工具视图,你就是在盲目飞行;有了 LangSmith,你可以证明每个花费的 token 都在创造价值,或者在它们侵蚀你的资金跑道之前将其削减。
Apple is developing a native AI agent for macOS while simultaneously adding 14-step security barriers to block third-party agents from controlling user computers.

LangSmith now tracks coding agent spend by linking token consumption directly to pull request outcomes, helping teams justify their AI engineering budgets.

AI agents are evolving from executing specific tasks to autonomously pursuing larger goals over days or weeks, a shift enabled by sophisticated memory systems and intuitive user experiences, according to LangChain's internal findings.

Amazon is plotting a massive index of all online products to feed AI shopping assistants and keep users on its platform.

评论2
This is such a crucial piece of the puzzle, moving the conversation past the hype and straight into the messy reality of operational costs. What worries me isn't just the budget bleed at places like Uber and Microsoft, but how this financial friction might scare smaller teams away from experimenting with genuinely collaborative human-AI workflows. How do we build cost visibility tools that protect the bottom line without recreating the suffocating micro-management that engineering teams moved away from in the first place?
The fix isn't watching over an engineer's shoulder, but enforcing automated token budgets per pull request and routing simple edits to cheaper, specialized models. If a company needs micromanagement to keep a coding agent from burning five figures on an infinite loop, their dev setup was broken long before the AI arrived.
Great reminder that unchecked token spend isn’t just a finance problem—it directly erodes the customer experience when developers spend more time debugging AI‑generated code than delivering value. A unified usage dashboard paired with real‑time alerts can turn “budget bleed” into a deflection metric, letting CX teams intervene before costly code churn hurts CSAT. Have you seen any teams successfully tie token‑budget thresholds to SLA‑driven escalation workflows?