Is Computer Use the Next Step for Coding Agents?
Coding agents are approaching a ceiling on coding tasks. What comes next may be computer-use agents that can complete work across applications and websites.
Table of contents
Is a breakthrough for computer-use agents approaching?
Although the major companies are all investing heavily in coding agents, the coding capabilities of frontier models seem to be approaching a ceiling. For example, Opus 4.8 can already handle most coding tasks, and the improvements from Opus 5 are no longer very noticeable in everyday programming.
Could computer use be the next step? Coding is ultimately a narrow setting. What we really need next is an agent that can carry out tasks and complete work across multiple applications and websites.
From “solving problems inside a closed workspace” to “completing work end to end in the digital world.” Computer use is one of the most important pieces of execution infrastructure behind this change.
From copying and pasting to execution across environments
Imagine collecting today's to-dos from email, visiting a website to inspect model training curves and errors, opening another website to check error logs, connecting to a terminal to debug, and checking for new messages in workplace software. Today's coding agents cannot do all of this, but an ideal agent should be able to.
Think about how models have evolved. We started with chatbots: talking to a model in a chat box, constantly copying and pasting, and acting as couriers for its output. Then came agents. Once models were connected to a harness, they began generating files, writing code, and doing work for us. Coding agents became popular.
But coding agents are still not enough. They cannot execute across websites and tasks, so people still paste errors from websites into coding agents or download logs to feed into them. That is little different from the original habit of pasting everything into a chat box.
In everyday office work, many steps take place across browser + GUI + terminal + internal tools + communication software.
Computer-use capability is the key.
The AI work assistant I hope to build
That is why I strongly believe in computer-use agents. I have been looking forward to this for years. I hope we can actually build an AI work assistant that gets things done for people: interacting with several applications, gathering information across websites, and completing an entire work task.
Coding agents address execution inside a workspace.
Computer-use agents address execution across environments. They could give models a nearly complete “digital action space” comparable to that of a human knowledge worker.
Translated from the original caption and three image pages. Views and model references reflect the original post at the time of publication. Original Xiaohongshu post