Realtime AI News
Google Unveils Agent Compute Engine, Giving Each AI Agent Its Own Isolated MicroVM
Google has introduced Agent Compute Engine, a managed Google Cloud service that runs AI agents in isolated microVMs and can checkpoint idle agents for later resumption. Credentials stay outside the sandbox and are injected by an external Agent Gateway, alongside an expanded set of agent governance controls.
Google has announced Agent Compute Engine, a managed service that runs AI agents inside isolated micro virtual machines and can checkpoint an agent's execution environment so it can be suspended and later resumed. The offering is aimed at Google Cloud customers and is positioned by the company as the compute foundation for what it calls the agentic era.
The most distinctive choice is what the design keeps out of the sandbox. API keys, OAuth refresh tokens and other credentials never enter the agent's guest environment; an external Agent Gateway holds them, authenticates each request and injects only the secrets a given tool call needs. In practice that means a prompt-injected or malfunctioning agent has no long-lived secret sitting inside the guest to steal.
Google frames the service as an answer to three requirements at global scale: keeping sandboxes free of exploitable credentials even when agents make mistakes, minimizing the cost and latency of resuming idle agents, and preserving persistent workspaces across sessions and operating system updates. The mechanism is copy-on-write memory sharing, which lets many agents run off a single operating system template, combined with fast checkpoint and restore of idle microVMs.
A companion deep dive by distinguished engineers Ramanathan Venkatapathy and Rohit Jnagal details a strict separation between the orchestration harness and the execution sandbox. The harness, described as the brain, runs entirely outside the guest; the sandbox, described as the hands, only executes explicit commands dispatched by the harness and streams output back. That boundary produces a deterministic audit trail correlating every harness-initiated tool call with the compute activity it actually caused.
Google sorts the workloads the platform is built for into four buckets: interactive consumer and enterprise agents that execute code, browse the web and take actions; persistent Linux workspaces for autonomous software engineering; large-scale parallel environments for model evaluation and reinforcement learning; and lightweight sandboxes for dynamic logic inside Google's own services.
Google says the platform already runs internally for Gemini, Gemini for Business, the agentic development platform Antigravity and model evaluation workloads. Alongside the managed service sits Agent Substrate, an open-source, Kubernetes-native agent runtime that manages sandbox lifecycles independently of the underlying infrastructure, suspending idle agents while preserving their state and resuming them on available workers.
On governance, Google is grouping the surrounding controls around centralized identity and policy enforcement: Agent Gateway for traffic control and policy, Agent Registry for tracking approved agents and tools, SPIFFE-based Agent Identity for authentication and auditing, and OpenTelemetry-compatible observability for monitoring model calls and tool invocations.
Agent Compute Engine is not yet available and pricing has not been announced; Google is targeting a preview in the coming months. Unlike the open-source Agent Substrate, which is portable across Kubernetes environments, the managed service is limited to Google Cloud. Competition is already present in the form of AWS Bedrock AgentCore and Azure Foundry Agent Service, which also offer managed, isolated execution environments for agents.
Outside analysis has flagged what the announcement does not say. AI specialist Harel Asaf notes that the deep dive gives no boot time, price or density figure for Agent Compute Engine itself, and that the numbers circulating in the announcement belong to Agent Substrate. The architecture and the governance story are clear; cost, performance and real availability are the questions a preview will have to answer.
Why it matters
As agents move from demos into production, the bottleneck shifts from model capability to isolation, credential handling and idle-time cost. Google is making the security boundary itself the selling point of a cloud service.
Nearby Updates
All11/22, 14:31
Amazon in Final-Stage Talks to Buy AI Startup Decart for About $7.5 Billion
Amazon is reportedly in the final stages of talks to acquire the AI startup Decart for about $7.5 billion, according to a Chinese tech news report published on November 22. Neither company has publicly confirmed the terms, so the transaction remains a late-stage negotiation rather than a completed acquisition.
11/22, 14:03
A社数学家用Claude又证明另一千禧难题,逼出一份写了12年的940页纯人类手稿
A社数学家用Claude又证明另一千禧难题,逼出一份写了12年的940页纯人类手稿.
11/22, 13:46
Google launches Gemini 4 Neon with faster performance and lower costs
Google has launched Gemini 4 Neon, a model it describes as faster and cheaper to run. The report gives no benchmark scores or pricing table, leaving the size of the improvement unconfirmed.
11/22, 13:38
Suno adds video generation to AI songs, powered by a Chinese open-source model
Suno has begun generating video for AI-created songs, extending its product from audio into visuals. Sohu reports the capability runs on a Chinese open-source model, though the specific model and rollout scope were not disclosed.