Giving an AI agent browser control unlocked tasks that previously had no API

The author initially kept an AI agent limited to the terminal, avoiding browser automation due to expected flakiness. A weekly report locked behind a login with no export option forced a change. Key challenges included detecting when pages were fully loaded using specific element waits rather than fixed delays, distinguishing real failures from slow loading, and clicking by accessibility role and name instead of unstable coordinates. The agent was also designed to stop and ask for human input at security prompts like two-factor authentication. The author found the real gain was enabling tasks previously thought to require manual work, such as reading dashboards and checking supplier portals, while still remaining cautious about partially working automation.