AI waits, its sandbox pauses: Google announces Agent Compute Engine
3-line summary
- Google announced Agent Compute Engine, with a virtual execution environment for each agent.
- When no commands or background tasks are active, idle environments preserve state and pause.
- A customer preview is targeted for the coming months; this is not general availability.
Google Cloud announced Agent Compute Engine on October 22. The service gives each agent a small virtual computer, or microVM. Google plans to offer infrastructure used by Gemini to external developers, targeting a preview in the coming months.
The technical explanation separates the harness, which calls the model and decides the next action, from the sandbox that runs code and browsers. The harness runs outside that sandbox. After a tool call, if no commands or background tasks remain active, the sandbox saves its state and becomes idle, then resumes for the next tool call.
Pausing the sandbox does not stop all model inference or the physical server. Neither official post specifies an exact preview date or price list. Fast resumption and efficiency are Google’s claims; this article does not report independent speed or cost measurements.
Why it matters
AI Ping sees sandbox idle time as another cost factor. Comparisons need published prices and workload patterns.
Who it's for
- Developers
- Map time spent executing tools, waiting for model responses and running background tasks before assessing infrastructure needs.
