Cloud sandboxes become the new normal for AI agents

The gist

Cloud sandboxes just became the gold standard for running scalable, secure AI agents—and AWS made it real.

What to know

Cloud Sandboxes: The New Baseline

AWS’s AgentCore V2 and Nitro Sandboxes cemented the shift away from risky local agent setups, setting a professional, standardized cloud foundation that redefines how AI agents are safely deployed at scale.

By November 2026, AWS’s AgentCore Runtime V2 and Nitro Sandboxes landed into a market that had already defined the problem: agents were too open-ended, too unpredictable, and too risky to keep running in ad hoc environments. Eye on AI had argued months earlier that sandboxed execution would become a new compute primitive because agents can spike CPU and memory use, perform potentially dangerous actions, and therefore need elastic, isolated infrastructure that can be spun up in large numbers, managed consistently, and discarded when work is done.

What made AWS’s November move decisive was not just better tooling, but the fact that it matched the operating model the rest of the conversation had converged on: agents need a standardized cloud surface, not a hobbyist machine under someone’s desk. Artificial Ignorance described the shift as agents moving off devices into persistent cloud environments, while The Verge reduced the requirement to a memorable baseline — “the right surface for that is 8 gigs of RAM and 8 gigs of storage in the cloud” — with the sandbox kept under professional control rather than on unmanaged consumer hardware.

Sources
Eye on AIArtificial IgnoranceThe Verge

MicroVMs Redefine Agent Hosting

Turning agent runtimes into isolated, elastic microVMs with scale-to-zero and persistent compute made autonomous, multi-turn AI sessions practical, affordable, and secure for the first time.

The category became mechanically credible when AWS turned the agent runtime itself into an isolated, elastic execution unit rather than a thin wrapper around containers. As CryptoRank.io noted, “Amazon released AgentCore Runtime V2 on September 18, 2026… V2 standardizes microVM-based isolated runtimes” and “introduces… scale-to-zero and persistent compute options,” alongside “consumption pricing ($0.1276 per vCPU-hour, $0.0169 per GB-hour),” which matters because the runtime now provides the persistence, resource elasticity, and enforcement boundary needed for multi-turn autonomous sessions instead of treating each invocation as a disposable job; as the article concludes, “Hosting, isolating, and scaling autonomous agents has transitioned.”

That architecture only works as a primitive for autonomous agents if startup and isolation are both strong enough, and AWS supplied both pieces. CryptoRank.io reported “a P75 cold start of 1.9 to 2.0 seconds for 200MB to 2GB images versus V1's 5.4 to 30 seconds,” adding that “this performance decouples image size from startup speed,” while AWS frames the sandbox mechanism as compatible with both managed container services and direct EC2 compute, and AWS’s Nitro Sandboxes can run on managed container services or EC2 and give each EKS pod or ECS task its own VM boundary on a shared host, providing hardware-level isolation and making secure, persistent cloud sandboxes practical rather than theoretical.

Sources

Snapshot Tech Slashes Latency

Snapshot-based environment restoration in AgentCore V2 delivers sub-two-second cold starts and tight resource control, transforming sandboxes from slow, costly outliers into agile, production-ready primitives.

What changes the economics is not just that sandboxes exist, but that they stop behaving like heavyweight, one-off environments. In the material here, that practical-economics case is supported less by explicit latency evidence than by stronger isolation paired with fast iterative feedback: the excerpt describes microVM isolation and “near real-time” browser updates, but it does not mention snapshot restoration or sub-two-second cold starts. By contrast, AWS said AgentCore Runtime V2 makes cloud sandboxes practical by cutting cold-start latency to around 2 seconds via snapshot-based environment preparation and restore: “AgentCore prepares your environment once, snapshots it, and restores the snapshot for every new instance,” keeping P75 cold starts at about 2 seconds for images from 200 MB up to 2 GB, where the original runtime rose from roughly 5.4 seconds to nearly 30 seconds as images grew.

That speed only matters economically because it arrives with tighter resource control and stronger isolation instead of looser, more expensive always-on setups. AWS says V2 requires explicit opt-in — “To opt in, you set platformVersion to V2 when you create or update a runtime” — and “your container must report healthy on /ping within 120 seconds,” while The Chenab Times showed the sandbox model extending to least-privilege access: developers “define a set of allowed ports (between 8000 and 12000) that the endpoint can reach, ensuring that only designated application ports are accessible,” with hostnames mapping to specific in-session ports.

Sources

Production-Scale Sandboxes Take Over

Platforms like DeepSeek DSec and Anthropic’s Claude Managed Agents prove that isolated cloud runtimes have become the backbone for both massive agent training and commercial deployment, replacing local execution across the industry.

The strongest evidence that cloud sandboxes are already the market’s default comes from builders operating at production scale, not from roadmap rhetoric. DeepSeek’s DSec platform, according to news.lavx.hu, “coordinates more than 380,000 concurrent sandboxes for agent training and evaluation” and “creates and manages millions of isolated execution environments for training and evaluating AI agents,” while also exposing containers, microVMs and full VMs behind a shared interface—evidence that serious agent systems are being designed around remote, isolated runtime fleets rather than local-only execution.

That pattern extends beyond training into commercial deployment and developer tooling. news.lavx.hu reported that DSec “handles more than 5,000 sandbox creations per second,” and that “A production-scale unit spans about 160 nodes and serves about 3 million sandboxes each day,” while Startup Fortune reported that “Agent37 now powers agents for more than 100 companies,” offering each customer a separate persistent cloud workspace instead of an ephemeral local run.

Anthropic’s November move shows the same convergence from another direction: even when customers want control, they are being handed infrastructure, not a laptop-bound agent package. On November 12, 2026, Anthropic’s ClaudeDevs account announced “self-hosted Claude Managed Agents in private beta,” stating “The server and the sandboxes stay with you, session-history database included,” and that “the entire service runs on your infrastructure,” with isolated sandbox containers tied to persistent sessions.

Sources

Get the stories behind the trends

Deep-dive reporting and the weekly brief, in your inbox.