At a large global manufacturer, the finance team spends weeks each quarter reconciling inventory against a mainframe-based system that no one else touches. The process lives on a PDF report, a legacy ERP screen, and a Citrix terminal. The team wrote an RPA bot to do it, but after every IT patch or UI refresh, the bot halts. Developers spend days rebuilding selectors, and the backlog grows. When a new module launched, the bot was dead in the water. This is the reality of RPA on mainframe and legacy systems: brittle bots, a rebuild treadmill, and SOPs that only humans can actually run.
Why RPA breaks here
Traditional RPA, whether from UiPath, Automation Anywhere, Blue Prism, or Power Automate, relies on selectors, xpaths, and object IDs to find the right element on a screen. When the application or the underlying UI framework changes, those identifiers shift. The bot fails, and a developer must rebuild it. In many organizations, the majority of RPA maintenance time is spent on this rebuild cycle rather than building new automations. Studies in the automation market show that as much as 60 to 70 percent of an RPA team’s effort is spent on maintenance and exception handling, not on new value. On legacy systems and environments like Citrix, where the UI is often virtualized and not exposed to the automation platform, selectors may not exist at all. Teams either abandon those processes or resort to brittle workarounds like image matching, which only works for static, unchanging screens. The result is a small set of highly stable processes kept alive, while the long tail of important but change-heavy work remains manual.
What changes with computer use agents
- Agents see the screen like a human and use the mouse and keyboard to interact, so they can work on any UI, including legacy screens and Citrix.
- They do not depend on brittle selectors or xpaths. When the UI changes, the agent continues to see and act on the updated elements.
- When they hit an exception or an unexpected state, they can reason about the situation and try an alternative path instead of halting.
- They follow SOPs written in plain English, which are already almost a prompt. No flowchart bots to design and babysit.
- They run across cloud VMs, a desktop app, and a /v1 computer use API, enabling parallel execution and integration into your existing toolchain.
Computer use agents replace brittle selectors with the ability to see and adapt, turning the rebuild treadmill into a one-time setup.
How to move without the risk
You do not have to rip out all RPA at once. A realistic path starts with a single high-impact process that is change-heavy and lives on a legacy or Citrix system. Pick a process where the business is already frustrated with manual effort and frequent bot failures. Run a pilot with a computer use agent: document the process as a clear SOP, deploy the agent to the same environment the RPA bot runs in, and compare uptime and effort. Measure how many changes you ship without bot disruption. Once you see sustained uptime and fewer rebuilds, expand to other processes on the same platform. Keep your stable, high-volume backend RPA automations where they excel, and layer in computer use agents for the changing and exception-heavy work. This phased approach lets you benefit from agents while staying confident in what you already have.
If you want to automate mainframe and legacy apps without a rebuild treadmill, the durable answer is computer use agents. Talk to the Coasty team to see how agents can take over the long tail of your processes at https://cal.com/coasty/15min .
Want to see this in action?
View Case Studies