Computer Use vs. Coding Agents: Which Is the Future?

GPT-6 Astra appears to lead by a wide margin on computer-use tasks in engineering. That is impressive—but I think it may be optimizing for the wrong interface.

Computer interfaces were designed for people

Graphical user interfaces primarily help people interact with computers in two types of scenarios:

1. Using functional software. A mechanical engineer designs a part in CAD, a designer edits an image in Photoshop, or a traveler books a trip on Expedia.
2. Consuming entertainment and social content. People watch short videos on TikTok, stream movies on Netflix or YouTube, and interact with others on social platforms such as Facebook.

The first category will change dramatically in a world of capable agents and powerful foundation models. Having an agent operate software by clicking through a GUI is often inefficient—and may eventually be unnecessary. For many functional tasks, an agent can work more directly and reliably through code or API calls.

The GUI will still matter, but its role will shift. It will become a layer for people to review results, inspect details, and verify that the agent did the right thing—not the primary interface through which the work gets done.

Why are frontier labs still focused on computer use?

My guess is that computer-use training data is comparatively easy to collect. A domain expert can complete a task while their screen and actions are recorded. The expert does not need to know how to program; they only need to understand the existing workflow.

Collecting high-quality coding-agent data for the same domain is much harder. It requires a rare combination of skills: someone who understands the domain, can write the necessary code, and knows how to evaluate both the implementation and its output.

This difference makes GUI interaction an accessible training target, even if it is not the most efficient long-term interface for agents.

The future

My prediction is that, within roughly a year, direct GUI operation will begin to fade from many functional workflows. Agents will increasingly perform the core work through code and APIs, while graphical interfaces remain important for human review, control, and verification.

Computer use is a useful bridge. It may not be the destination.