Google Cloud says its new GKE Agent Sandbox starts AI training sandboxes 10 to 45 times faster

What Google Cloud announced

Google Cloud made the announcement in a blog post dated September 30, 2026. It covers three things: a version of GKE Agent Sandbox optimized for reinforcement learning, the Agent Sandbox RL orchestration SDK (a software toolkit for managing many sandboxes at once), and native integrations with common RL "gyms" and tools, which are ready-made training environments. Google Cloud says all three are now generally available. The post was written by product manager Tinsley Shi and staff software engineer and technical lead Tomer Glottmann.

Google Cloud describes the problem this way. When an AI agent is trained with RL, a model running on GPUs (costly AI chips) produces actions such as pieces of code. Those actions are then run in a sandbox, an isolated computer environment on ordinary CPUs, so the result can be scored as feedback. At large scale, the GPUs sit idle while they wait. They wait for sandboxes to cold start (boot from scratch), for many multi-gigabyte container images (packaged copies of the software each task needs) to download, and for scheduling queues to clear.

Read the full description

Sources are collected, independently checked, then reviewed by Jev.

GKE is Google Cloud's service for running Kubernetes, a widely used open system for running software across many machines. Google Cloud describes GKE Agent Sandbox as an open Kubernetes building block for running agents securely. It includes a built-in SandboxWarmPool, a "warm pool" of environments started in advance so they are ready to use. It also supports snapshots, pause and resume.

• Orchestration SDK: Google Cloud says it offers an asynchronous Python API, so researchers do not have to write Kubernetes configuration files (YAML).
• Tool integrations: Google Cloud lists Gymnasium, NVIDIA NeMo Gym and OpenHands.
• Controller: Google Cloud says Agent Sandbox Controller v1.0.0 limits how quickly the warm pool is refilled. This keeps etcd and the API server stable. Both are core Kubernetes components that track and manage the cluster.
• Customer: Google Cloud says the component already powers the agentic RL training infrastructure at Mistral AI, which it calls a frontier AI lab.

The benchmark figures Google Cloud published

Google Cloud says it tested on a cluster of 10 machines (nodes) running gVisor sandboxes, with GKE Image Streaming turned on. The tests used two workloads: a SWE-bench coding environment with 500 images, and an R2E collection with 4,578 images. Between 500 and 18,312 tasks ran at the same time. The R2E-Gym-Subset dataset page on Hugging Face lists 4,578 rows, which matches the size Google Cloud gives. That page says nothing about performance.

Google Cloud credits two changes for the speed-up. First, the software images are made ready ahead of time rather than while training waits. Second, the pace at which the cluster's control plane handles requests is regulated. The control plane is the part of Kubernetes that manages everything else.

A pod is the basic unit Kubernetes runs, and a rollout is one attempt by the agent at a task. Google Cloud says fewer pods are created because the SDK reuses the same pods between rollouts instead of deleting and rebuilding them. Google Cloud also says its benchmarking tools and load tests come with the agent-sandbox-rl example, so teams can repeat the comparison on their own clusters.

Why this matters

This is a change to the behind-the-scenes infrastructure that AI companies use. Many AI agents are now trained to write code and use tools, and that training means running attempts again and again in securely isolated environments. Google Cloud points out a bottleneck in this process: in a synchronous RL training step, training cannot continue until the slowest sandbox in the batch is ready. So slow-starting sandboxes leave expensive GPUs waiting.

If Google Cloud's claims hold, AI labs could do more agent training and testing with less wasted computing power. That could indirectly speed up work on products such as AI coding assistants. Google Cloud also quotes Mistral AI research engineer Jean-Malo Delignon as saying they can handle spikes of more than 30,000 sandboxes on a single cluster.

Trade-offs and limitations

Google Cloud acknowledges a trade-off: the approach spends cheaper CPU and background cluster time preparing environments in advance, so the GPUs are not left idle. It also says the SDK counts every failed sandbox and marks it for retry rather than quietly dropping it, because untracked failures can skew the rewards the model learns from. The update does not directly change any apps that everyday users have. It mainly affects research teams and startups that train AI agents at large scale.

Browse the latest news in this topic

• Tech news