Google Cloud made reinforcement-learning Agent Sandbox generally available on Google Kubernetes Engine, with 1–9-second time-to-first-command.

Google Cloud announced general availability of Agent Sandbox optimized for reinforcement learning (RL) on Google Kubernetes Engine (GKE), alongside an RL orchestration software development kit (SDK) and native integrations for RL environments and evaluation harnesses. The product targets workloads in which a model generates actions on GPUs and executes them in isolated CPU environments to obtain reward signals.

Google reports 1–9 seconds to first command, compared with 45–85 seconds previously, describing the improvement as 10–45 times faster. It also reports reducing worst-case sandbox wait times from 7.5 minutes to under 10 seconds; these figures concern sandbox startup and waiting, not GPU arithmetic throughput.

The SDK recycles existing pods between rollouts instead of repeatedly deleting and recreating them. Google says this produces three times fewer pod creations, reducing control-plane churn during bursts involving thousands of sandbox requests.

The engineering work used demanding agent benchmarks, including SWE-bench-style workloads, to expose problems such as scheduling backlogs, large image pulls, and control-plane timeouts. These are Google’s reported engineering and performance results; the supplied excerpt does not quantify customer cost savings or end-to-end training improvements.