Google Puts Gemini Agents on Empty Machines and Pauses Them While Waiting for Humans

Google is preparing to open up a runtime configuration, and this configuration is currently what runs Gemini's own agents. There are three key details: each agent gets an isolated machine; no credentials are stored on the machine; the agent pauses when waiting for human input. A preview version is expected in the coming months.

This news comes from a tweet by William Geronco. The original text is short, with the information concentrated on how it runs. Original link

The difference between an agent and a chat box is that it executes tasks. Execution touches files, networks, and third-party services, and it also requires permissions. A common approach is to stuff credentials into environment variables or give the agent an account that can access many resources. It's convenient, but the boundaries are fuzzy. Google's configuration does the opposite: the agent has its own machine, and the machine contains no credentials. The security boundary falls on the runtime environment; prompts can't govern that.

'Pausing while waiting for a human' is also worth a look. The agent stops midway and waits for a human to confirm or add information, which means it cannot continue acting while the person is away. The cost is that the workflow breaks, and efficiency won't look as good as full automation. The benefit is that the control points are clear: a human can intervene at key moments, the machine state is preserved, and it continues running after recovery.

The tweet's author offers a judgment: this is how agents should run. That judgment has a real-world point. Rather than giving the model more permissions, give it fewer credentials. An agent's capabilities can be very strong, but the runtime environment must keep it contained.

What to watch next are the engineering details: how credentials are injected, how machine lifecycles are managed, whether pause and resume are stable, and whether audit records are complete. The preview isn't here yet, and these questions determine whether it can go from a demo to an everyday tool. For AI practitioners with a development background, this is a change at the runtime layer; for ordinary users, it can be understood this way: an agent is not a longer chat history—it is a machine that moves, and someone has to watch it.