GPT-6 Astra in Action: One Vague Prompt, Full Feature Built and Tested

OpenAI's next-generation model GPT-6 Astra is being pushed to its limits by the AI team at fintech company Ramp. In a conversation with Ramp's head of AI, we got a clear picture of how Astra performs in real-world workflows — less like a tool that needs step-by-step instructions, and more like a colleague who understands your intent and can drive tasks forward on their own.

A Trusted Colleague, Not a Tool That Needs a Hundred Lines of Instructions

Ramp's head of AI had a straightforward first impression of Astra: it's a colleague you can trust. You don't need to hand it a hundred lines of instructions — just give it a high-level goal and the outcome you want to see, then come back in 10 to 15 minutes to review.

"Astra is really good at taking these vague requests and knowing exactly what output I want, then returning screenshots of the changes," he said. That capability made him comfortable handing Astra increasingly ambiguous tasks.

The real "wow" moment came from the speed at which Astra moved tasks forward without any human intervention. Its communication style was refreshingly concise — no unnecessary filler, which is invaluable in fast-paced work environments where you need results, not pleasantries.

Computer Use: The Coding Harness Can Trigger Computer Operations

The most technically notable detail from the conversation was that Astra's coding harness can trigger computer use — meaning the model doesn't just write code, it can interact with interfaces just like a human would to verify results.

Ramp's AI lead walked through a specific example: he submitted a feature request — "implement routing strategies controls by API key granularity" — with no additional context or explanation.

What was surprising was that Astra figured out the navigation path on its own: it needed to click on an existing API key, then click the routing strategy option to surface the relevant interface. He didn't have to explain "click this tab, on this page, this route, this modal" — Astra "just knew."

What is Computer Use? Computer Use refers to an AI model's ability to directly control a mouse, keyboard, and screen — browsing the web, clicking buttons, and filling out forms just like a human user. Anthropic was the first to publicly demonstrate this capability in its Claude models in late 2024, with OpenAI subsequently incorporating it into its model roadmap. The core breakthrough is that the model no longer just generates text or code — it can perceive the current state of an interface through screenshots and plan its next actions accordingly. This is fundamentally different from traditional RPA (Robotic Process Automation) tools, which rely on pre-recorded fixed scripts and break whenever the UI changes. LLM-based Computer Use can understand the semantics of an interface and adapt accordingly. Astra's "coding harness triggering computer use" means the model can seamlessly switch between writing code and interacting with a browser to verify results — forming a closed development-and-validation loop.

12 Minutes to Build, 15 Minutes to Self-Verify

In this feature implementation case, Astra worked for about 12 minutes to complete the development, then spent another 15 minutes verifying its own changes. It precisely identified the models and settings that needed to be modified.

Even more unexpected: Astra verified the changes on mobile — a step the user never asked for. Ramp's AI lead believes computer use will increasingly become an important way to validate low-risk changes: the model doesn't just complete the task, it proactively ensures results hold across multiple scenarios.

Next Step: Putting Computer Use in Customers' Hands

Looking ahead, Ramp is most excited about putting computer use capabilities directly in customers' hands. One of Ramp's core value propositions is automating and simplifying accounting workflows, and computer use is central to delivering on that.

Currently, Ramp uses computer use technology to fetch receipts from merchant websites on behalf of customers in the background — but it requires a significant amount of manual prompting to ensure the model knows where to navigate on each merchant's site.

Ramp's AI lead is eager to test Astra against longer-tail merchant sites to see how it handles them. He believes Astra has a better general understanding of how websites work, which could reduce the need for that tedious manual guidance — a critical step toward truly scaling AI capabilities to end users at scale.

A Closer Look: Receipt Fetching as a Long-Tail Automation Challenge The "receipt fetching" use case Ramp describes is a classic long-tail automation problem: after employees make purchases on hundreds of different merchant platforms — Amazon, Airbnb, Uber, and more — they need to log into each platform and download e-invoices for expense reimbursement and tax compliance. Because billing page structures vary wildly across merchants, traditional automation scripts must be written separately for each site, making maintenance extremely costly. Ramp's current approach relies on heavy "manual prompting" to tell the model how to navigate each site — essentially hardcoding human knowledge. If Astra can leverage its general understanding of web structures to navigate merchant sites autonomously, it would dramatically reduce the human effort required to onboard new merchants, making this class of capability genuinely viable at scale.

What This Conversation Reveals About the Evolution of AI Agents

This Ramp × GPT-6 Astra conversation — sourced from YouTube and presented from the vendor's perspective — nonetheless illuminates several key trends in AI Agent development: a shift from "precise instructions" to "intent understanding," an expansion from "writing code" to "operating interfaces and self-verifying," and a progression from internal tooling to customer-facing automation services.

When a model can receive ambiguous goals, autonomously plan its operation path, and verify results across multiple environments, the boundary of human-AI collaboration is being fundamentally redefined. For businesses like Ramp that depend on large volumes of repetitive web operations, the maturation of these capabilities means more long-tail, tedious tasks can be delegated to AI.

The Technical Underpinnings of "Intent Over Instructions" The shift from "precise instructions" to "intent understanding" described here corresponds to a core research direction in AI: raising the abstraction level of task specification. Early AI assistants required users to describe every step in a structured, unambiguous way. The current generation of Agent models attempts to receive high-level natural language goals and then autonomously decompose subtasks, infer missing information, and formulate an execution plan. This capability is supported by several technical pillars: stronger contextual understanding and long-horizon reasoning, the combination of tool use (function calling/tool use) with multimodal perception, and task planning strategies optimized through reinforcement learning. Astra's unprompted decision to verify on mobile exemplifies this "goal-oriented rather than instruction-oriented" autonomy — the model has internalized the implicit engineering norm that "feature verification should cover multiple platforms," rather than merely executing the literal request.