Tech • AI • Robotics • Game
One Prompt. A Feature Built and Tested. | Ramp × GPT-6 Astra
6/10TL;DR
Ramp is testing Astra as an AI coworker for software development and web-based automation, with early results showing it can complete vague coding tasks, validate low-risk changes, and potentially improve how the company retrieves customer receipts from merchant websites.
KEY POINTS
AI used as a high-level collaborator
At Ramp, Astra is being used less like a traditional coding assistant and more like a trusted coworker that can act on broad goals. Instead of requiring long step-by-step prompts, it can take a loosely defined request, work independently for several minutes, and return concrete results such as interface changes and screenshots.
Focus on faster software development
The main use case so far is speeding up coding workflows and improving implementation quality. The system is described as effective in day-to-day engineering tasks because it is concise, avoids unnecessary back-and-forth, and can infer details about navigation, tabs, routes and modal windows without explicit instructions.
Computer-use capability is central
A key feature is computer use, which allows the coding setup to interact directly with software interfaces. That means the model is not limited to generating code; it can also click through products, inspect settings, and verify that requested changes appear correctly in the user interface.
Vague feature request completed end to end
In one example, the prompt was simply to implement routing strategies controls by API key granularity. From that short request, Astra identified that the relevant controls were accessible by selecting an existing API key and opening routing strategies, then spent about 12 minutes making changes and roughly another 15 minutes validating them.
Validation went beyond the original request
The tool did more than just implement the feature. It also verified the result in mobile view, even though that check was not explicitly requested. That kind of autonomous validation is seen as especially useful for confirming low-risk product changes before human review.
Early impression centers on autonomy
The strongest initial reaction was how quickly the system made meaningful progress without further user input. Rather than needing constant supervision, it could be left to work and checked on 10 to 15 minutes later, with enough context awareness to understand the intended outcome from a short prompt.
Potential customer-facing applications
Beyond internal engineering, Ramp is exploring how similar computer-use capabilities could benefit customers directly. One area of interest is extending the company’s automation around accounting-related tasks so that AI can simplify more of the operational work users would otherwise handle themselves.
Receipt retrieval is a practical test case
A current example is background receipt fetching from merchant websites. Ramp already uses computer-use tools for this, but the process still requires substantial manual prompting so the model knows where to navigate on each merchant’s site. The hope is that Astra can generalize better across a long tail of merchants by recognizing common website patterns with less hand-holding.
CONCLUSION
The testing suggests that Astra could become useful not only for writing code but also for navigating software and websites with limited supervision. If that autonomy holds up across more workflows and more merchants, it could make both internal development and customer-facing automation significantly more efficient.
Ask a question