Desktop AI Agents May Be the New Excel

I am starting to think that the most useful AI agents will not take the form we first imagine. We talk a lot about chatbots, voice assistants, and autonomous agents able to execute entire missions. But in my daily use, the agents that feel most interesting are desktop agents. They are in the right place: directly inside the environment where work happens.

A useful agent does not only need to answer a question. It needs to access files, understand how a project is organized, modify code, run a command, read an error, compare two versions, and resume a task after a failure. These actions are ordinary, but they represent a large part of real work. And that work does not happen in a chat window, unless you enjoy copy-pasting all day. It happens in a folder, a terminal, an editor, a browser, a business application, or a set of local files.

This is why tools like Claude Code, OpenAI Codex, and OpenCode seem more important than they first appear. They can be described as development assistants, which is true, but too narrow. Their value is not only that they help write code. Their value is that they put AI inside the work context, with the ability to observe, suggest, modify, and verify. The agent becomes less of an interlocutor you ask questions to and more of a tool you work with on material that is already there.

This is where the comparison with Microsoft Excel becomes useful. Excel did not win because it was the best database tool, the best programming language, or the best reporting system. It won because it allowed people close to the problem to build good-enough solutions quickly, without going through a full software production process. Excel is rarely the perfect tool, but it was available, flexible, understandable, and immediately actionable.

Desktop agents could occupy a similar place. They will not replace business software, technical teams, or robust systems. But they can become the intermediate layer that helps people turn intent into action more easily. Cleaning a folder, automating a repetitive task, adapting a script, generating a small tool, understanding why something does not work, producing a usable first versionWhy desktop AI agents could become the next productivity layer by helping people automate, adapt, and build tools directly where work happens.: these are ordinary needs, but their accumulation represents a huge share of daily productivity.

What seems underestimated is that an agent’s usefulness depends much less on its appearance than on its location. An agent in a chat interface remains separated from the work. You have to bring it context, copy elements, rephrase the task, retrieve the answer, and then go back somewhere else to act. A desktop agent reduces that distance. It can work in the same space as the user, on the same files, with the same constraints. That proximity changes the nature of usage.

We often imagine agents as highly autonomous systems, able to go off and complete complex tasks on their own. Some use cases may go in that direction. But I am not sure that is the most important starting point. The truly useful agents may first be tools that stay very close to the user, help with concrete tasks, and have a limited but sufficient level of autonomy to save time.

From this perspective, desktop agents look less like chatbots and more like a new interface layer between the user and the work environment. The web browser structured a large part of modern computing. Chat was the first natural interface for generative AI. But for agents, the most durable format may be closer to the operating system than to messaging.

That is the sense in which desktop agents may be the new Excel. They serve a comparable function: giving people close to the problem the ability to build, adapt, and automate their own tools. Excel made modeling and automation accessible to millions of people. Desktop agents could do the same for a much wider part of digital work.