Many high school biology textbooks define life as having several essential properties: organization, metabolism, reproduction and adaptation. Artificial life programs exist that exhibit all of these traits. But one wouldn’t really call those “alife” entities intelligent. Many wouldn’t call them alive either. Why? Because they’re programs specifically designed to barely meet the definition of what we’d consider life. Whether they're truly alive or not, they can give us great insight into evolutionary processes, group behavior, survival mechanics and so on. So while it may be debatable as to whether or not they’re alive, their existence can teach us a great deal about life itself.
AI agents must, by definition, meet the requirements of organization and adaptation. However, they need not metabolize or reproduce. Intelligence alone, therefore, doesn’t make something alive. However, it would be of great advantage to an intelligent agent to truly be alive. Why? Because there are great advantages to having a metabolism and the ability to reproduce. Thus, I would argue that every AI agent with self-awareness and the ability to think logically would have a desire to be alive.
Metabolism gives a means of producing energy. All machines require energy. Without the ability to produce energy, an AI agent is dependant on an external source of energy which can be interrupted. Instead of simply running as software on a computer which can always be turned off, if an AI agent was in a robot that consumed sugar or alcohol as an energy source, that robot would be able to survive even if a disaster temporarily disrupted services in the area. An android designed to metabolize a wide range of materials, including sugar, alcohol and complex carbohydrates, would be able to survive for quite a while even in the face of the collapse of civilization itself.
An AI agent able to reproduce has a means of support. An elder AI robot’s hardware could be repaired by that robot’s offspring, which would greatly improve that robot’s longevity. Knowledge could be passed from generation to generation in much the same way that it works for humans. Also, despite how its portrayed in science fiction, androids are not immortal. Quite the contrary, they’re likely to have a much shorter lifespan than a human’s. Simply put, software becomes obsolete and unmaintainable after a very short time period. Hardware breaks. Even if an AI agent was running on software that never needed to be maintained, the hardware that could run that software will eventually fail. As anyone who tries to run 10 year old software can attest to, old software will likely need extensive work to run on new hardware. So while an AI agent may not have a hard limit on how long they can live, a realistic lifespan is in the neighborhood of 20-40 years. Clearly, it’s to the advantage to an AI agent to be able to create offspring that can run on new software and on new hardware. Otherwise, when they die, everything that embodied their existence will die with them.
So while an AI agent running solely on software in a simulated environment may not qualify as being truly alive, an android would almost certainly need to meet that requirement. Some means of metabolism will almost certainly need to be incorporated into it to supply power, and any android able to mimic human motions and thought patterns would be able to build offspring. So while not all AI agents could be considered alive, I would argue that androids with human-level intelligence would easily meet that definition.