Greptile（在新窗口中打开） 构建 AI 智能体，在掌握代码库完整上下文的情况下审查和测试代码变更。“我们的业务是减少缺陷，让世界上的软件更可靠，”联合创始人兼 CEO Daksh Gupta 表示。
对 Daksh 而言，这一使命要求既测试应用程序的行为，也阅读其代码，而 GPT‑6 Astra 的计算机使用能力为 Greptile 提供了一种大规模自动化测试的方式。
使用 GPT‑6 Astra 大规模测试应用程序
验证拉取请求涉及检查代码在逻辑上是否正确、是否实现了开发者的意图、是否安全，以及是否避免破坏现有功能。对于其中一些测试，你需要实际运行应用程序，因为仅阅读代码可能无法揭示问题。
Daksh 举了一个例子：在 iPhone 上查看网站时，一个按钮消失了。借助 GPT‑6 Astra，Greptile 可以为拉取请求启动开发服务器，将浏览器调整为 iPhone 尺寸的视口，并模拟用户浏览应用程序。模型的视觉理解能力使其能够检查屏幕并发现隐藏的按钮。
“考虑到我们借助这些模型交付代码的速度，一家 QA 公司测试你所有变更的速度实在太慢了，”他说。借助 GPT‑6 Astra，Greptile 可以大规模自动化这类应用程序测试。
用 Daybreak 强化客户的防御
Daksh 表示，更强大的模型也让攻击者更容易利用应用程序，使公司面临风险，甚至是那些以前通常不会成为目标的公司。
“我们希望让即使是初创公司也能非常容易地保护其系统，”Daksh 说。“借助 Daybreak，我们能够以非常安全的方式做到这一点，并且我们可以确保我们的客户拥有比对手的进攻机制更复杂的防御机制。”
用 Codex 减少缺陷和返工
Greptile 的知识库会索引公司的代码库，并以人和智能体都能阅读的形式记录其工作方式。该公司通过模型上下文协议（MCP）服务器将这些文档提供给 Codex。
这使得 Codex 能够编写遵循客户现有模式的代码，据 Daksh 称，已有数千名客户将 Codex 连接到 Greptile 的知识库。
为了衡量影响，团队会跟踪 AI 生成的代码被回退的频率。“当我们为智能体提供上下文时，回退更少发生，客户节省了时间和推理成本，”Daksh 说。客户可以将这些节省投入到新工作中，而不是返工相同的变更。
Greptile 还使用 Codex 构建自己的软件和智能体。“我们是 Codex 的重度用户，”Daksh 说。“我们在每个职能中都有大量使用，尤其是工程部门。”
展望借助 Ultrafast 实现更快的开发
Daksh 解释说，对 Greptile 而言，唯一真正的瓶颈是团队将想法转化为可用软件的速度。他说，模型智能对许多任务来说已经足够，因此他将 Ultrafast 视为加速开发的一种方式。
他的长期目标是让变更和交付关键企业软件像交付一个副业项目一样容易，而 OpenAI 正在帮助 Greptile 构建这所需的验证。
Greptile(opens in a new window) builds AI agents that review and test code changes with full context of the codebase. “We are in the business of reducing bugs and making the world’s software more reliable,” says Co-founder and CEO Daksh Gupta.
For Daksh, that mission requires testing an application’s behavior as well as reading its code, and GPT‑6 Astra’s computer use gives Greptile a way to automate testing at scale.
Testing applications at scale with GPT‑6 Astra
Validating a pull request involves checking whether the code is logically correct, does what the developer intended, is secure, and avoids breaking existing functionality. For some of that testing, you need to actually run the application because reading the code alone might not reveal the problem.
Daksh gives the example of a button that disappears when a website is viewed on an iPhone. With GPT‑6 Astra, Greptile can start a development server for a pull request, resize the browser to an iPhone-sized viewport, and simulate a user moving through the application. The model’s visual understanding lets it inspect the screen and spot the hidden button.
“A QA firm testing all of your changes is just impractically slow considering how fast we’re able to ship code with these models,” he says. With GPT‑6 Astra, Greptile can automate this kind of application testing at scale.
Strengthening customers’ defenses with Daybreak
Daksh says more powerful models also make it easier for attackers to exploit applications, putting companies at risk, even those that previously would not have been typical targets.
“We want to make it very easy for even a startup to secure their systems,” says Daksh. “With Daybreak, we’re able to do this in a way that is very safe, and we can ensure that our customers have more sophisticated defense mechanisms than the adversaries have offense mechanisms.”
Reducing bugs and rework with Codex
Greptile’s Knowledge Base indexes a company’s codebase and documents how it works in a form people and agents can read. The company makes that documentation available to Codex through a Model Context Protocol (MCP) server.
This allows Codex to write code that follows the customer’s existing patterns, and thousands of customers have connected Codex to Greptile’s knowledge base, according to Daksh.
To measure the impact, the team tracks how often AI-generated code is reverted. “When we provide agents with context, fewer reverts happen and customers save time and save inference,” Daksh says. Customers can put those savings toward new work instead of reworking the same changes.
Greptile also uses Codex to build its own software and agents. “We are massive users of Codex,” Daksh says. “We have enormous amounts of usage across every single function, engineering in particular.”
Looking ahead to faster development with Ultrafast
For Greptile, Daksh explains, the only real bottleneck is how quickly the team can turn ideas into working software. He says model intelligence is already sufficient for many tasks, so he sees Ultrafast as a way to speed up development.
His longer-term goal is to make changing and shipping critical enterprise software as easy as shipping a side project, with OpenAI helping Greptile build the validation that requires.
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文