Code review startup Greptile is extending its AI agents beyond reading code to running applications, using OpenAI's GPT-6 Astra and its computer-use capability to test software the way a user would. Co-founder and CEO Daksh Gupta said the company's mission is reducing bugs and making software more reliable, and that some problems only appear when an app is running. He gave the example of a button that disappears when a website is viewed on an iPhone. Greptile's agent can start a development server for a pull request, resize the browser to phone size, step through the app, and use the model's visual understanding to spot the missing element. Gupta argued that human quality assurance teams cannot keep pace with how quickly companies now ship code. Greptile is also using Daybreak to strengthen customers' security defences. Gupta said more capable models make it easier for attackers to exploit applications, including at companies that were not previously typical targets, and that the aim is to let even small startups secure their systems. On the coding side, Greptile's Knowledge Base indexes a company's codebase and documents how it works in a form that people and AI agents can read. That documentation is available to OpenAI's Codex through a Model Context Protocol server, and Greptile says thousands of customers have connected Codex to it. The company tracks how often AI-generated code is reverted as a measure of impact, and Gupta said better context means fewer reverts and savings in time and inference costs. Greptile is also a heavy user of Codex internally. The article is a company case study published by OpenAI, so the performance and cost claims are self-reported and have not been independently verified. For Australian businesses and professionals, this changes little unless they build or ship software. For engineering teams, automated end-to-end testing against staging environments is worth piloting, but costs scale with model usage. Any testing that touches customer personal information should be assessed against the Privacy Act 1988 (Cth) and internal security policies before rollout. Automated review and testing should supplement human review rather than replace it.