Self-Hosted AI Workspaces - Native Desktop Agents vs Browser Sandboxes

A thread about self-hosted agentic AI workspaces with subagent loops raised an important architectural question: should AI workspaces run in a browser sandbox or as native desktop agents?

The answer depends entirely on what you are trying to automate.

Browser-Based Workspaces

Tools that run AI agents in a browser environment - whether cloud-hosted or self-hosted - give the agent access to web applications through a sandboxed browser instance. The agent can navigate websites, fill forms, click buttons, and extract data from web pages.

This works well for tasks that live entirely on the web. If your workflow is "check this dashboard, extract these numbers, paste them into this spreadsheet, send this email" and all of those are web apps, a browser-based agent handles it.

The sandbox is both a feature and a limitation. Your data stays contained, which is good for security. But the agent cannot reach anything outside the browser - your local files, your native apps, your system preferences, your terminal.

Native Desktop Agents

A native agent runs on your operating system and interacts with applications through accessibility APIs. On macOS, this means the agent can control any app that has a UI - not just web apps, but Finder, Terminal, Mail, Calendar, Xcode, Slack, and everything else.

The agent sees your actual desktop state. It knows which windows are open, what text is selected, what notifications have appeared. It can drag files between apps, copy data from one native app to another, and chain together workflows that cross the boundary between web and desktop.

Where the Lines Blur

The interesting middle ground is agents that do both. A native macOS agent can control a browser just as easily as any other app - it has access to browser tabs, forms, and content through the same accessibility framework. But it can also step outside the browser to work with local files, run terminal commands, or interact with native applications.

Self-hosted browser workspaces are a good fit for teams that need controlled, reproducible environments. Native agents are better for individual productivity where the goal is automating your personal workflow across all your tools - not just the ones that run in a browser.

The trajectory of the space suggests native agents will absorb the browser-based use cases over time, since they can do everything a browser agent does plus more.

Fazm is an open source macOS AI agent. Open source on GitHub.