Greptile extends bug detection with GPT-6 Astra workflow testing

Greptile adds application testing to AI code review

Greptile is using GPT-6 Astra to extend bug detection beyond code review by testing how a running application behaves. The startup builds AI agents that review and test code changes with context from a customer’s codebase, and it says computer use enables automated user-workflow testing at scale.

Co-founder and CEO Daksh Gupta said the company’s aim is to reduce bugs and improve software reliability. In Greptile’s approach, validating a pull request means more than deciding whether its code is logically correct: it also involves checking whether the change does what the developer intended, remains secure and does not break existing functionality.

Why a running application can reveal defects

Some failures cannot be detected by reading code alone. Gupta cited a website button that disappears when the site 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 can then inspect what appears on screen and identify the hidden button. Gupta said that having a QA firm test every change would be impractically slow given the pace at which teams can ship code with models. Greptile positions automated computer-use testing as a way to perform this type of application validation at scale.

Security and codebase context are part of the workflow

Greptile also uses Daybreak to strengthen customer defenses. Gupta said that more capable models can make it easier for attackers to exploit applications, including at companies that may not previously have appeared to be typical targets. He said Greptile wants to make it easier for startups to secure their systems while providing defense mechanisms that are more sophisticated than adversaries’ offensive capabilities.

For coding work, Greptile’s Knowledge Base indexes a company’s codebase and documents how it works in a form that people and agents can use. The company exposes that documentation to Codex through a Model Context Protocol, or MCP, server. Gupta said thousands of customers have connected Codex to the Greptile knowledge base.

That context is intended to help Codex produce code consistent with a customer’s existing patterns. Greptile tracks how often AI-generated code is reverted and reports that fewer reverts occur when agents receive context, saving customers time and inference that can be redirected to new work rather than repeated changes.

What development teams can take from the example

Greptile also uses Codex internally across its software and agent development. Gupta sees Ultrafast as a way to accelerate the conversion of ideas into working software, while describing the ability to move from idea to implementation as the central bottleneck. The practical implication for businesses is to pair AI-assisted coding with runtime workflow validation, security testing and codebase-specific context before treating a change as ready to ship.