总部位于纽约、利用AI代理查明公司软件为何崩溃的初创公司Traversal周二宣布，已在B轮融资中筹集了1.2亿美元。此轮交易对其估值为10亿美元。
Traversal, a New York startup whose AI agents find out why company software breaks, has raised $120 million in a Series B round. The deal values it at $1 billion, the company announced on Tuesday.
此轮融资由凯鹏华盈（Kleiner Perkins）、红点创投（Redpoint Ventures）和科斯拉创投（Khosla Ventures）共同领投。红杉资本（Sequoia Capital）、电池风投（Battery Ventures）和美国运通风投（Amex Ventures）也参与了投资，其中凯鹏华盈和红杉资本均增持了股份。至此，Traversal的总融资额已达到1.73亿美元。
Kleiner Perkins, Redpoint Ventures and Khosla Ventures co-led the round. Sequoia Capital, Battery Ventures and Amex Ventures also took part, and Kleiner Perkins and Sequoia both increased their stakes. Traversal has now raised $173 million in total.
其观点是，AI编写代码的速度超过了人类维持其运行的速度。联合创始人兼首席执行官阿尼什·阿加瓦尔（Anish Agarwal）对《福布斯》表示，这种新软件也更难修复。
Its argument is that AI is writing code faster than people can keep it running. Co-founder and chief executive Anish Agarwal told Forbes that this new software is also harder to fix.
阿加瓦尔对《福布斯》说：“它们的崩溃方式有着根本上的不同。”
“They break in fundamentally new ways,” Agarwal told Forbes.
Traversal的业务是网站可靠性工程，这是一项保持网站和应用程序正常运行的工作。当服务宕机时，工程师会通过日志、指标和代码来追踪故障。Traversal的代理能够进行这种搜索、测试可能的成因并提出修复建议。
What Traversal does Site reliability engineering is the job of keeping websites and apps running. When a service goes down, engineers trace the fault through logs, metrics and code. Traversal’s agents do that search, test possible causes and suggest a fix.
该公司表示，可观测性工具是为人工阅读仪表盘而构建的，而不是为代理设计的。其生产世界模型（Production World Model）将遥测、代码和部署历史转化为AI可以读取的地图。随后，其因果搜索引擎（Causal Search Engine）会同时测试数千种可能的成因。
The company says observability tools were built for people reading dashboards, not for agents. Its Production World Model turns telemetry, code and deployment history into a map an AI can read. Its Causal Search Engine then tests thousands of possible causes at once.
Traversal工作线程（Traversal Workers）会在警报触发后立即开始调查。该公司表示，Traversal MCP会将系统学到的有关生产环境的知识传递给AI编码代理，以便它们避免引发下一次故障。
Traversal Workers start an investigation as soon as an alert fires. Traversal MCP passes what the system learns about production to AI coding agents, so they can avoid causing the next incident, the company says.
阿加瓦尔向《福布斯》讲述了一家银行的案例：当客户无法使用服务时，该银行召集了数千名工程师进入“作战室”。原因是他系统深处的一个响应迟缓的数据库。他说，几个月后当类似问题再次出现时，Traversal的代理在五分钟内就找到了问题。
Agarwal told Forbes about one bank that gathered thousands of engineers in a “war room” when customers could not. The cause was a slow database deep in its systems. When a similar problem came back months later, Traversal’s agents found it in five minutes, he said.
该公司表示，Nubank、Harvey、DigitalOcean、百事公司（PepsiCo）和美国运通（American Express）都是Traversal的客户。《福布斯》报道称，黑石集团（Blackstone）也是其客户。
Customers and growth Nubank, Harvey, DigitalOcean, PepsiCo and American Express use Traversal, the company said. Forbes reported that Blackstone is also a customer.
Traversal表示，自今年年初以来，其年度经常性收入已经增长了九倍。阿加瓦尔拒绝向《福布斯》透露具体的收入数字。该公司称，客户在上一季度进行了超过15万次调查。今年，每月的调查次数从513次激增至近7.3万次。
Traversal says its annual recurring revenue has grown nine times since the start of the year. Agarwal declined to give Forbes revenue figures. Customers ran more than 150,000 investigations last quarter, the company said. Monthly investigations rose from 513 to nearly 73,000 this year.
该公司称，目前约有三分之二的使用场景已经超出了查找宕机原因的范畴。客户还将该产品用于警报分类、发布风险和安全性。
About two thirds of its use now goes beyond finding the cause of an outage. Customers also use it for alert triage, release risk and security, according to the company.
Nubank创始人兼首席执行官大卫·韦莱斯（David Vélez）表示：“在我们推出Traversal五周后，它就已经成为我们值班工程师处理故障工作的一部分。”
“Five weeks into our Traversal rollout, and it’s already becoming part of how our on-call engineers work incidents,” said David Vélez, founder and chief executive of Nubank.
Traversal与康奈尔科技学院（Cornell Tech）和哥伦比亚大学共同构建了一个名为ORCA-bench的基准测试。它测试通用AI代理在1079个任务中的表现，这些任务要求它们找到故障的根本原因。
A benchmark for AI on call Traversal built a benchmark called ORCA-bench with Cornell Tech and Columbia University. It tests general AI agents on 1,079 tasks where they must find the root cause of an incident.
Traversal表示，表现最好的五个前沿代理在处理中等难度故障时的准确率为25.3%。当故障报告含糊不清时，准确率降至10%。幻觉率达到了40%。
The best of five frontier agents got 25.3% of medium-difficulty incidents right, Traversal said. Accuracy fell to 10% when incident reports were vague. Hallucination rates reached 40%.
阿加瓦尔说：“网站可靠性工程是应用AI中最棘手的问题之一，因为每家公司的软件环境都各不相同，而且宕机每年给公司造成高达6000亿美元的损失，事关重大。”
“Site reliability engineering is one of the hardest problems in applied AI because every company’s software environment is different, and with downtime costing companies as much as $600 billion a year, the stakes are enormous,” said Agarwal.
几位创始人在麻省理工学院（MIT）获得了博士学位，拉杰·阿加瓦尔（Raj Agrawal）和拉兹·德维维迪（Raaz Dwivedi）的博士学位则是在加州大学伯克利分校获得的。艾哈迈德·隆（Ahmed Lone）此前曾在城堡证券（Citadel Securities）从事量化交易。阿加瓦尔和德维维迪保留了他们在哥伦比亚大学和康奈尔科技学院的教职。
The founders Agarwal and Raj Agrawal earned their PhDs at MIT, and Raaz Dwivedi earned his at UC Berkeley. Ahmed Lone came from quant trading at Citadel Securities. Agarwal and Dwivedi have kept their faculty posts at Columbia University and Cornell Tech.
Traversal于2025年6月结束隐蔽状态，获得了来自红杉资本和凯鹏华盈的4800万美元资金。美国运通风投于2025年12月进行了投资。这家初创公司计划在研究、工程和市场开拓团队中进行招聘。
Traversal came out of stealth in June 2025 with $48 million from Sequoia and Kleiner Perkins. Amex Ventures invested in December 2025. The startup plans to hire across research, engineering and go-to-market teams.
凯鹏华盈合伙人马蒙·哈米德（Mamoon Hamid）对《福布斯》表示，AI“让网络可靠性问题变得异常严峻”。
AI has “created acuteness around the problem of cyber reliability,” Kleiner Perkins partner Mamoon Hamid told Forbes.