An agent waiting for a model still has a working directory, unfinished edits and a process to resume. Turning off its compute is easy. Keeping its place is the harder part.
Google’s November 5 Agent Substrate announcement describes a runtime that separates those jobs. An Actor represents a session’s filesystem, environment, memory and network rules. A pool of warm Workers supplies the compute when the session needs to run. After a turn, the runtime can save the Actor’s state and release its worker slot.
Google says the project has reached version 1.0 and announces the donation of Agent Substrate and gVisor to the Cloud Native Computing Foundation in Sandbox status. These are Google’s statements about the project’s release and governance, rather than an independent assessment of production readiness.
The mechanism matters for agents that spend long periods waiting between short bursts of work. Instead of keeping a dedicated Kubernetes Pod occupied for each session, a platform can move saved sessions through shared execution capacity. Google reports restoration in under 500 milliseconds and describes gVisor’s user-space interception of system calls as part of the isolation boundary. The announcement supplies no independent measurement of that restoration time.
For a platform team, the useful experiment is a recovery test with recognizable work inside it. Pause an agent after an uncommitted file edit, restore it on another worker, and inspect the file, running process and network access. Then repeat with a worker failure. A fast wakeup is useful only if the resumed session preserves the right state and the intended permissions.
Measure storage and restoration overhead beside the compute saved during idle time. The decision belongs to the complete session lifecycle, including the moment someone has to recover work that stopped midway.