AI agents should not just write code
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,…
AI agents should transcend their current role as mere code-writing entities. Current workflows for coding agents are limited to editing files, executing tests, possibly initiating a development server, and reporting back. This is insufficient for web applications. A human developer not only reviews diffs but also interacts with the application.
They navigate the app, click through the user journey, identify issues such as incorrect tabs being open, buttons not functioning, unexpected page changes, or the browser still pointing to yesterday's backend. Crucially, the browser state becomes a coordination aspect. When multiple agents operate simultaneously, browser state introduces another layer of coordination complexity.
Without proper isolation, browser tabs may point to the wrong workspace, agents may click in another agent's application, authentication states may leak between branches, screenshots may fail to match the running process, and the developer faces challenges in determining which agent is performing which task. Consequently, Agent-Up now treats the browser as an integral part of the workspace runtime.
Each workspace is assigned its own browser session, application tabs, port mappings, logs, and runtime state. Agents interact with this browser through MCP (Message Control Protocol), enabling them to navigate the app, inspect the page, click links and buttons, wait for specific text or selectors, capture screenshots, and record browser activity into an audit history.
The key point is not merely that the agent can click; it is that the click is associated with the correct workspace. If an agent clicks within its application, Agent-Up can adjust the visible application tab for that workspace, shift the agent mouse representation, exhibit the click animation, and maintain the alignment between the developer's desktop view and the agent's actual actions.
The review process transforms. Instead of perusing multiple terminal logs and attempting to decipher what transpired, you can observe four agents concurrently exercising four isolated versions of an application. One agent might scrutinize the login flow, another examines pricing, another traverses the checkout process, and another validates returns and fulfillment.
While sharing the same machine and repository, each agent operates within distinct worktrees, runtime environments, and browser sessions. The browser has become an essential component of the feedback loop. This aspect of agentic development is often underestimated: the browser is not merely an output surface. It is an integral part of the feedback loop.
For web applications, an agent devoid of browser utilization is missing the very product it aims to support. Git isolates the source code, while Agent-Up isolates the runtime environment, and MCP provides agents with control over both. This enables easier supervision of parallel agent work because one is no longer merely reviewing generated code; they are witnessing each agent operate its own version of the product.
This is precisely what the demo aimed to illustrate. Agent-Up can be downloaded from https://agent-up.themassiveone.net/ and its open-source repository is available at https://github.com/agent-up-oss/agent-up/.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
