A team of researchers built EnvOS, a framework for testing how AI agents that control software cope when that software changes mid-task. The core finding: agents fail far more often on dynamic environments than static tests reveal. The failure mode is consistent. The agent carries an old picture of the screen and acts on it, producing errors that compound. The researchers tested two ways to help agents recover from a changed interface. One approach sent the agent a reminder that things might have changed. The other refreshed the agent's view of the current screen before it acted. The interface-level refresh produced better outcomes. The reminder, delivered as text, did not close the gap. The agent needed to see the updated state, not read about it. For teams building or evaluating agents that control software, EnvOS offers a structured approach to certification. Construct tasks around observed failure patterns. Require evidence that the agent handled each one. Score against that evidence. The framework treats a test suite as something that can go stale too, and builds renewal into the process.