MarbleOS tests a visible desktop workspace for AI agent work

Why it matters

AI agents are gaining the ability to browse, edit files and run recurring tasks. MarbleOS is testing whether the winning interface will organize that work as visible objects instead of messages.

Miguel Acevedo, the engineer behind MarbleOS, is showing a macOS workspace that keeps an AI agent's files, tools, tasks and outputs visible while it works.

The interface is Acevedo's answer to a problem that has followed agents from research demos into everyday software: a chat transcript is a poor filing system. An agent can browse websites, assemble a spreadsheet and schedule a recurring task, yet the user is often left scrolling through messages to reconstruct what happened.

MarbleOS gives that work a desktop-style surface. MarbleOS's demo page describes a workspace with visible files, tools, tasks and outputs. MarbleOS currently has a public-access waitlist and a macOS DMG download page.

Acevedo has been building tools around overlooked workflow friction since he was a computer science student at the University of California, Davis. In 2017, he and Sergei Chestakov built Schedule Helper, a Chrome extension that flagged class conflicts and brought professor ratings into the university's scheduling system. The California Aggie reported that the pair completed the initial project in a few weeks and approached 1,000 downloads around its release. Acevedo's LinkedIn profile later claimed 13,000 active users.

He followed that project with Marble LeetCode Tutor, a browser extension that guides programmers through coding problems. The Chrome Web Store identifies Acevedo as its developer and shows 23 ratings, with the most recent listed update dated June 4, 2025. The products differ, but the pattern is consistent: Acevedo tends to place guidance and automation inside the workspace where a user is already trying to finish a task.

The workspace is the product

MarbleOS is competing at the interface layer. Its demo library contains 16 clips, each under two minutes, covering lead research, client-meeting preparation, candidate review, outreach drafting, competitor research and recurring morning briefs.

Those examples point toward founders, sales teams, recruiters, client-service operators and small businesses. MarbleOS has not presented that list as a formal customer definition; it is visible in the work Acevedo chose to demonstrate. These are jobs that produce artifacts worth keeping: spreadsheets, briefs, task lists and research notes.

That output-first design matters once an agent handles a sequence of actions rather than answering a single question. If a run produces a spreadsheet, a research brief and a weekly routine, the user needs to know which files are current, which sources informed them and what will happen again without another prompt. A running conversation can describe those objects. A workspace can organize them.

MarbleOS also emphasizes review. The product page says users can hand off the browsing and assembly work, then return to a first pass. That keeps the human in the operating loop without requiring the user to watch every click. The design still has to prove that its visible objects accurately reflect the agent's actions.

The permissions behind that work will matter as much as the layout. MarbleOS's privacy policy, last updated April 26, 2026, is part of that early trust test. Users will judge the working product by how clearly it communicates access, approvals and completed actions.

The large labs already have execution

Acevedo is making this interface bet after agent capabilities moved into the largest AI assistants. OpenAI's ChatGPT agent release notes describe an agent that can complete complex online tasks using tools including a virtual browser and terminal. Anthropic has also described computer-use and agentic workflows in which Claude can point, click and interact with desktop environments with user permission.

That creates a difficult distribution problem for a small product. OpenAI and Anthropic can place execution inside assistants that users already know. MarbleOS has to make the workspace itself valuable enough to justify another application.

The same pressure creates Acevedo's opening. As assistants take on longer jobs, their interfaces have to represent plans, permissions, intermediate work, reusable files and scheduled activity. A single input box handles requests well. It becomes strained when the assistant starts behaving like an operator with several pieces of work in progress.

MarbleOS is betting that users will want the agent's activity laid out like work on a computer rather than compressed into a record of conversation. Acevedo's earlier products suggest an instinct for removing friction from familiar tasks without asking users to learn an elaborate system first. Schedule Helper surfaced conflicts where students chose classes. The LeetCode tutor placed instruction beside the coding problem. MarbleOS applies the same approach to agents by putting their work where users can see and reuse it.

The early test will be repeat use. One-off research demos can make nearly any agent interface look effective. Recurring briefs, follow-up routines and evolving spreadsheets create a harder standard: MarbleOS must preserve state, expose mistakes and make it clear what the agent will do next. If Acevedo can make those behaviors legible, the workspace may become the reason users choose MarbleOS.