Google launches GKE Agent Sandbox for reinforcement learning

Google has rolled out GKE Agent Sandbox, a Kubernetes-based system designed specifically for agentic reinforcement learning (RL) workloads. The platform promises to cut environment startup time from up to 85 seconds down to as little as 1 second, reducing GPU idle time. Worst-case wait times for sandboxes have dropped from 7.5 minutes to under 10 seconds, and control-plane churn is down threefold, thanks to in-place pod recycling. The system is available now with SDK support for RL researchers.

• Cuts rollout time from 85 seconds to 1–9 seconds
• Worst-case wait for sandboxes now under 10 seconds
• Pod recycling cuts control-plane churn by three times
• Supports large-scale, parallel AI agent training
• SDK and popular RL tool integration included