Early this year, Jacob Andreou had a crucial goal. The Microsoft executive set out to order a McDonald’s cheeseburger to his apartment across from the company’s Silicon Valley campus using a homegrown AI tool. Andreou and his team had worked nonstop to get the tool, named Copilot Tasks, up and running. If the AI agent inside Tasks could autonomously order the burger and get it to Andreou’s home, it would prove the tool worked as designed. By the time Andreou reached his apartment, the burger had...