They should be able to use the application they changed.
That sounds obvious, but most coding agent workflows still stop at editing files, running tests, maybe starting a dev server, and reporting back. For web apps, that is not enough.
A human developer does not only inspect diffs. They open the app. They click through the flow. They notice when the wrong tab is open, when a button does nothing, when the page changed unexpectedly, or when the browser is still pointing at yesterday's backend.
AI agents need that same feedback loop.
Once multiple agents are running in parallel, browser state becomes another coordination problem. Without isolation, browser tabs point to the wrong workspace, one agent clicks in another agent's app, authentication state leaks between branches, screenshots no longer match the running process, and the developer cannot tell which agent is doing what.
This is why Agent-Up now treats the browser as part of the workspace runtime.
Each workspace gets its own browser session, application tabs, port mappings, logs, and runtime state. Agents interact with that browser through MCP, so they can navigate to the app, inspect the page, click links and buttons, wait for text or selectors, take screenshots, and record browser activity into the audit history.
The important part is not just that the agent can click. It is that the click belongs to the correct workspace.
When an agent clicks inside its app, Agent-Up can switch the visible application tab for that workspace, move the agent mouse representation, show the click animation, and keep the developer's desktop view aligned with what the agent is actually doing.
Instead of reading four terminal logs and guessing what happened, you can watch four agents exercise four isolated versions of an app at the same time.
One agent might test the login flow. Another checks pricing. Another walks through checkout. Another validates returns and fulfillment.
Same machine. Same repo.
But different worktrees, different runtime environments, and different browser sessions.
This is the part of agentic development I think people underestimate: the browser is not just an output surface.
It is part of the feedback loop.
For web applications, an agent that cannot use the browser is missing the place where the product actually exists. Git isolates the source. Agent-Up isolates the runtime. MCP gives agents control over both.
That makes parallel agent work easier to supervise because you are no longer only reviewing generated code. You are watching each agent operate its own version of the product.
That is what the demo was built to show.
Download Agent-Up:
[https://agent-up.themassiveone.net/](https://agent-up.themassiveone.net/)
Open source repo: