--- title: Accelerate agentic RL with GKE Agent Sandbox | daily.dev description: Google Cloud announced general availability of GKE Agent Sandbox optimized for reinforcement learning (RL), alongside an Agent Sandbox RL orchestration SDK and... canonical: https://daily.dev/posts/accelerate-agentic-rl-with-gke-agent-sandbox-v8mk9efre twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Accelerate agentic RL with GKE Agent Sandbox | daily.dev og:description: Google Cloud announced general availability of GKE Agent Sandbox optimized for reinforcement learning (RL), alongside an Agent Sandbox RL orchestration SDK and... og:url: https://daily.dev/posts/accelerate-agentic-rl-with-gke-agent-sandbox-v8mk9efre og:image: https://api.daily.dev/og/posts/v8MK9EFre.png og:image:alt: Accelerate agentic RL with GKE Agent Sandbox og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Accelerate agentic RL with GKE Agent Sandbox **[Google Cloud](https://daily.dev/sources/gcp)** · 8 min read · 1 upvotes · 1 comments ## Summary Google Cloud announced general availability of GKE Agent Sandbox optimized for reinforcement learning (RL), alongside an Agent Sandbox RL orchestration SDK and native integrations for RL gyms and harnesses like Gymnasium, NVIDIA NeMo Gym, and OpenHands. The offering targets bottlenecks in agentic RL infrastructure: slow sandbox cold-starts, massive image cardinality, and control-plane saturation from bursty pod churn. Benchmarks show 10x-45x faster time-to-first-command (1-9 seconds vs 45-85 seconds), tail latency reduced from 7.5 minutes to under 10 seconds, and 3x fewer pod creations via an in-place pod recycling strategy. Mistral AI is cited as a production user, reporting handling spikes of over 30,000 sandboxes on a single cluster. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is GKE Agent Sandbox optimized for RL and what problem does it solve? GKE Agent Sandbox optimized for RL is a Kubernetes primitive from Google Cloud, now generally available, that eliminates sandbox cold-start bottlenecks in agentic reinforcement learning. It uses built-in SandboxWarmPool capabilities combined with GKE Image Streaming to cut time-to-first-command from 45-85 seconds down to 1-9 seconds, keeping expensive GPU/TPU accelerators from sitting idle during RL training rollouts. _Teams scaling agentic RL pipelines can track infrastructure releases like this via daily.dev._ ### How much does the Agent Sandbox RL orchestration SDK reduce Kubernetes control-plane churn during RL training? The SDK's in-place pod recycling strategy reduces pod creations by 3.1x during multi-trajectory rollouts by running an in-pod git reset and repository checkout between rollout episodes instead of deleting and recreating pods. This keeps the Kubernetes control plane flat under burst loads that would otherwise trigger unbounded garbage-collection loops at scales like 18,312 simultaneous sandboxes. _Engineers tuning Kubernetes for bursty AI workloads can follow infrastructure deep dives like this on daily.dev._ ### How is Mistral AI using GKE Agent Sandbox for reinforcement learning? Mistral AI uses GKE's Agent Sandbox for RL to orchestrate hundreds of thousands of secure execution environments across clusters, handling spikes of over 30,000 sandboxes on a single cluster to accelerate model training and iteration cycles. Research Engineer Jean-Malo Delignon credited it as the reliable foundation enabling instant scaling to unpredictable RL demand. _Researchers evaluating RL infrastructure choices can compare real-world deployments like this on daily.dev._ ## Community discussion Top comments from developers on daily.dev. **@nahum** · 0 upvotes > 3x fewer pod creations is the number I'd watch. our worst day at close to two thousand gVisor sandboxes per node group came from the delete side: one lost signal kept a pod Terminating for 5 days. does in-place recycling mean a sandbox never goes through teardown at all? ## Similar posts on daily.dev - [Google Announces GKE Agent Sandbox and Hypercluster at Next '26, Positioning Kubernetes as AI Agent](https://daily.dev/posts/google-announces-gke-agent-sandbox-and-hypercluster-at-next-26-positioning-kubernetes-as-ai-agent-cfmrtqf9a) · InfoQ · 0 upvotes · 0 comments - [What’s new in GKE at Next ‘26](https://daily.dev/posts/what-s-new-in-gke-at-next-26-082dwhthn) · Google Cloud · 0 upvotes · 0 comments - [Google Cloud: A Deep Dive into GKE Sandbox for Agents](https://daily.dev/posts/google-cloud-a-deep-dive-into-gke-sandbox-for-agents-keoxzofzg) · The New Stack · 3 upvotes · 0 comments - [Building a Cluster-Aware AI Agent with Kubernetes, Argo CD, and GitOps](https://daily.dev/posts/building-a-cluster-aware-ai-agent-with-kubernetes-argo-cd-and-gitops-6omlndjcx) · CNCF · 2 upvotes · 0 comments --- Tags: [#kubernetes](https://daily.dev/tags/kubernetes), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning), [#ai-infrastructure](https://daily.dev/tags/ai-infrastructure), [#gke](https://daily.dev/tags/gke) [View this post on daily.dev](https://daily.dev/posts/accelerate-agentic-rl-with-gke-agent-sandbox-v8mk9efre) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Accelerate agentic RL with GKE Agent Sandbox","url":"https://daily.dev/posts/accelerate-agentic-rl-with-gke-agent-sandbox-v8mk9efre","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/accelerate-agentic-rl-with-gke-agent-sandbox-v8mk9efre"},"datePublished":"2026-09-29T17:05:35.583Z","dateModified":"2026-09-29T17:05:59.349Z","description":"Google Cloud announced general availability of GKE Agent Sandbox optimized for reinforcement learning (RL), alongside an Agent Sandbox RL orchestration SDK and...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/71a924f2b121958fb9f73a189cb1663d?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/71a924f2b121958fb9f73a189cb1663d?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Google Cloud","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Google Cloud","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/gcp","url":"https://daily.dev/sources/gcp"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/accelerate-agentic-rl-with-gke-agent-sandbox-v8mk9efre","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"kubernetes,reinforcement-learning,ai-infrastructure,gke","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Google Cloud","item":"https://daily.dev/sources/gcp"},{"@type":"ListItem","position":3,"name":"Accelerate agentic RL with GKE Agent Sandbox"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/accelerate-agentic-rl-with-gke-agent-sandbox-v8mk9efre","comment":[{"@type":"Comment","text":"3x fewer pod creations is the number I’d watch. our worst day at close to two thousand gVisor sandboxes per node group came from the delete side: one lost signal kept a pod Terminating for 5 days. does in-place recycling mean a sandbox never goes through teardown at all?","datePublished":"2026-10-02T08:56:53.788Z","url":"https://daily.dev/posts/v8MK9EFre#c-nOCkfiCnE","author":{"@type":"Person","name":"Nahum Litvin","url":"https://daily.dev/nahum","image":"https://lh3.googleusercontent.com/a-/AOh14GhvJ2y-9R45d5ee78ZdAo8l7J3t4G8fINkx5PEqBfA=s100"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/accelerate-agentic-rl-with-gke-agent-sandbox-v8mk9efre#faq","mainEntity":[{"@type":"Question","name":"What is GKE Agent Sandbox optimized for RL and what problem does it solve?","acceptedAnswer":{"@type":"Answer","text":"GKE Agent Sandbox optimized for RL is a Kubernetes primitive from Google Cloud, now generally available, that eliminates sandbox cold-start bottlenecks in agentic reinforcement learning. It uses built-in SandboxWarmPool capabilities combined with GKE Image Streaming to cut time-to-first-command from 45-85 seconds down to 1-9 seconds, keeping expensive GPU/TPU accelerators from sitting idle during RL training rollouts. Teams scaling agentic RL pipelines can track infrastructure releases like this via daily.dev."}},{"@type":"Question","name":"How much does the Agent Sandbox RL orchestration SDK reduce Kubernetes control-plane churn during RL training?","acceptedAnswer":{"@type":"Answer","text":"The SDK's in-place pod recycling strategy reduces pod creations by 3.1x during multi-trajectory rollouts by running an in-pod git reset and repository checkout between rollout episodes instead of deleting and recreating pods. This keeps the Kubernetes control plane flat under burst loads that would otherwise trigger unbounded garbage-collection loops at scales like 18,312 simultaneous sandboxes. Engineers tuning Kubernetes for bursty AI workloads can follow infrastructure deep dives like this on daily.dev."}},{"@type":"Question","name":"How is Mistral AI using GKE Agent Sandbox for reinforcement learning?","acceptedAnswer":{"@type":"Answer","text":"Mistral AI uses GKE's Agent Sandbox for RL to orchestrate hundreds of thousands of secure execution environments across clusters, handling spikes of over 30,000 sandboxes on a single cluster to accelerate model training and iteration cycles. Research Engineer Jean-Malo Delignon credited it as the reliable foundation enabling instant scaling to unpredictable RL demand. Researchers evaluating RL infrastructure choices can compare real-world deployments like this on daily.dev."}}]} ```