Google Cloud said on 22 October 2026 that it is launching Agent Compute Engine, “a purpose-built compute service that already powers Gemini as an agent for work tasks,” announcing it at the AGNTCon + MCPCon conference. “It will be available in preview in the coming months,” Google wrote. In the same post, Google said its open-source Agent Substrate runtime “delivers as much as 10x higher compute density compared to traditional compute” for AI agents.
In plain words, an AI agent that writes code, runs tools or drives a browser needs a computer to act on. Google’s product gives each agent its own small, isolated virtual machine, a microVM, that starts quickly and can be paused while the agent is idle. The post also expands Google’s rules and controls for agents, which it calls agent governance, to agents running outside Google Cloud.
Business Pill · THE MODEL AND THE HARNESS
Google’s benchmark figures for Agent Substrate, in three steps. Source: Google Cloud blog, 22 October 2026; Google’s own figures.
A short Business Engineer explainer on the difference between an AI model and the harness around it. It teaches the general idea only and says nothing about any company in this piece.
The key insight: As we read it, Google is treating the computer an agent runs on as its own product: per its posts, each agent gets an isolated microVM that is paused while it waits, and the density figure Google gives, for Agent Substrate, comes from packing such idle agents onto one machine.
Update (23 Oct): the idle-time bill, and who already cut it.InfoWorld reported on 23 October that Google is betting idle infrastructure costs could become a major barrier to enterprise adoption of AI agents. Google’s own technical deep dive puts the design rule plainly: “Agents spend most of their time waiting. An idle agent should cost almost nothing and wake up faster than a person notices.”
Google is not first to stop paying for the wait. Amit Kumar Jena, AI development head at IT consulting firm Kanerika, told InfoWorld: “AWS has run a microVM per session in AgentCore since 2025 and stops billing CPU while the agent waits. Microsoft’s hosted agents in Foundry suspend between turns and keep the filesystem.” Google’s step beyond, as Jena described it, is checkpointing that keeps memory and execution state, so an agent resumes where it left off. Manoj Chandra Jha, principal analyst at Nord-IQ Research, told InfoWorld that continuously active, compute-intensive agents are likely to see smaller gains, and that the service is not yet in preview.
The catch is in the agent code, Jena said: “If the pause lands between an API call and its confirmation, the agent wakes up unsure whether the payment went through or the commit landed, and it tries again. Every external action has to be safe to repeat. Most agent code today isn’t written that way.” He also flagged storage, since each paused agent keeps a disk. His test for buyers: “CIOs should ideally count how many agents hold state across more than a few minutes of idle time.” Idle compute is the same cost line Nebius went after when it bought Inferize.
Update (23 Oct): a startup selling the same thing raised at $1 billion. Daytona, which sells cloud computers built for AI agents, announced on 20 October a $60M Series B at a $1B valuation, led by FirstMark Growth, bringing its total funding to $91M. In its own words, an agent’s computer “has to start in milliseconds, pause with its state intact, fork into parallel copies, and disappear when the job is done,” across Linux, Windows, macOS and GPU sandboxes.
The figures are the company’s: Daytona says its revenue has grown “approximately 25×” since January, with average month-over-month net revenue retention of “approximately 135%,” and that one customer, Writer, “provisioned 300,000 sandboxes in a single month.” Tech.eu reported that Daytona provides infrastructure to Anthropic, OpenAI, NVIDIA, Cursor and Cognition. Pausing an agent with its state kept is the same designgoalGoogle sets for Agent Compute Engine above.
Google framed the product around its own use. “Scaling Gemini to planet scale taught us firsthand that securely running autonomous agents is fundamentally an infrastructure challenge,” the post opens. It says Agent Compute Engine gives every Gemini agent its own secure computer and that customers will soon be able to use the service for their own agents.
Speed and cost.Google says the service uses “copy-on-write memory sharing to run multiple agents on a single operating system template” and saves capacity by “pausing idle microVMs via fast checkpoint and restore.”
Isolation. Each microVM runs behind a hardware-isolated hypervisor boundary on Titanium, Google’s offload processing unit, according to the post.
Memory between sessions. Each microVM has its own Hyperdisk storage, so an agent’s workspace and context persist, Google says.
“Agent Compute Engine microVMs give us the responsiveness, hardware isolation, and workload density needed to deliver interactive agent sandboxes to our users around the world,” said Gautham Thambidorai, VP and Google Fellow for Gemini, in the post.
Google says its open-source Agent Substrate runtime delivers as much as 10x higher compute density for AI agents than traditional compute, and its 15 September post says it can pack over 1,000 dormant agents per host. The figures are Google’s own.
What Was Posted on X
Announced at AGNTCon + MCPCon: the new Agent Compute Engine plus expanded Gemini for business agent governance deliver the control you need to run, govern, and scale your agents.
Learn more about Agent Compute Engine and expanded agent governance → https://goo.gle/4yFN3yT
Google Cloud’s own post on 22 October, at 15:06 UTC, announced the launch at AGNTCon + MCPCon: the new Agent Compute Engine plus expanded Gemini for business agent governance, which it says deliver the control companies need to run, govern and scale their agents. The post links to the launch blog this piece is built on.
GOOGLE 🔥: Managed Agents are coming to Google AI Studio soon.
Users will be able to “Inspect and manage their autonomous AI agents, their tools, environments, and triggers” from a dedicated Agents hub.
Earlier this week, Google announced its new Agent Compute Engine for GCP, on top of which this solution is likely being built.
— 🚨 AI News | TestingCatalog (@testingcatalog) Oct 25 13:32
TestingCatalog (@testingcatalog) reported on 25 Oct that Managed Agents are coming to Google AI Studio, with a dedicated Agents hub where users can “Inspect and manage their autonomous AI agents, their tools, environments, and triggers.” TestingCatalog says it is likely built on Agent Compute Engine. This piece treats it as an unconfirmed leak.
The Data Point: How Many Agents per Machine
Google’s own deep dive sets out the design premise: “Agents spend most of their time waiting. An idle agent should cost almost nothing and wake up faster than a person notices.” The figures Google gives are all for Agent Substrate, the open-source runtime for Google Kubernetes Engine that it presented alongside the new service; for Agent Compute Engine itself, the posts give no density figure and no price.
What Google measured or promised
Figure, as Google states it
Where Google said it
Agent density on Agent Substrate
As much as 10x higher compute density compared to traditional compute
Launch post, 22 Oct 2026
Idle agents packed on one machine
Over 1,000 dormant agents per host
Agent Substrate on GKE post, 15 Sep 2026
Benchmark size
200,000 agents
Launch post, 22 Oct 2026
Activation rate in that benchmark
500 activations per second
Launch post, 22 Oct 2026
Suspend and resume latency in that benchmark
Less than 500 ms
Launch post, 22 Oct 2026
Agent Compute Engine availability
Preview in the coming months; no price published
Launch post and deep dive, 22 Oct 2026
Sources: Google Cloud blog posts as dated. The density and benchmark figures are Google’s own and describe Agent Substrate, the open-source runtime Google presented alongside Agent Compute Engine; we have not seen independent tests of them.
The 10x figure is a ratio Google gives for Agent Substrate against what it calls traditional compute, not a measured comparison with any named rival.
Governance for Agents Outside Google
The second half of the announcement extends Google’s agent controls: Agent Gateway for traffic and policy, Agent Registry as a list of approved agents and tools, Agent Identity to give each agent its own ID, and observability for every model and tool call. Google says it is extending governance to coding agents running on company laptops: the post lists “Extends governance to coding agents like Claude Code and Antigravity CLI on managed devices with an OS-level sandbox and local governance services,” in preview soon.
Coming this quarter in preview, Google says, token management will let companies “attribute spend, enforce budget guardrails” across agents and models, with model routing to send requests to the right model. Ping Identity, SailPoint and Saviynt are named as identity partners.
Who Is Using It
Google names Nous Research as an Agent Substrate user and quotes its chief business officer, Hervé Bizira: “Agent Substrate addresses both at the platform layer in a way that also preserves valuable compute resources.” In its 15 September post, Google says Nous Research’s Hermes agent “is currently ranked the #1 AI agent globally by OpenRouter usage”. The launch post adds that Agent Substrate “works with any agent framework or harness, including Claude Code, OpenClaw, and Hermes.” L’Oréal is quoted as a user of the governance tools on Cloud Run.
Connecting the Coverage
Our Claude-on-Google-Cloud update reported that Anthropic said on 22 October that Claude Models on Google Cloud are generally available and that Claude usage draws down a customer’s existing Google Cloud commitment. What’s new: by our reading, Google is positioning its own compute and controls to run agents whichever model sits inside them.
Our Agent Skills piece reported that Anthropic said on 22 October 2026 that it is donating Agent Skills, the open standard that packages instructions and scripts an agent can load for a task, to the Agentic AI Foundation under the Linux Foundation. Our bank-agents piece reported, citing Axios on 22 October 2026, that major US banks are detecting unauthorized third-party AI agents inside customer accounts, and that four of the five banks Axios names are blocking or restricting outside agents, by our count. Google’s post describes a different approach to agent access: a separate identity for every agent.
Our Nadella harness piece reported that Microsoft’s CEO wrote on 21 October 2026 that Microsoft runs a system where the harness is separate from the model, so that “models become interchangeable.” Google’s post makes a related claim for its runtime, saying it works with any harness.
The Structural Read
As we read it, the design starts from a usage pattern Google states in its deep dive: agents spend most of their time waiting, so an idle agent should cost almost nothing.
As we read it, the governance half reaches beyond Google’s own runtimes: the post says it will extend controls to agents on other clouds, on premises and on managed laptops, naming Claude Code and Antigravity CLI.
As we read it, the commercial terms are still missing: Google gives no price and no preview date for Agent Compute Engine, and its density figures describe the open-source Agent Substrate runtime.
Google Cloud, Agent Compute Engine deep dive, 22 October 2026
“An idle agent should cost almost nothing and wake up faster than a person notices”
Three Implications
FOR COMPANIES RUNNING AGENTS Google says Agent Compute Engine already runs Gemini’s agents and will reach customers in preview in the coming months; Agent Substrate is open source and runs on any Kubernetes cluster, according to the post.
FOR SECURITY TEAMS Google says each agent gets its own identity based on the SPIFFE standard, with governance extending to coding agents on managed devices in preview soon.
FOR FINANCE TEAMS Google says token management will let companies attribute spend and enforce budget guardrails across agents and models, in preview this quarter.
The framework’s core line: “The surface, not the model, governs autonomy.”
As we read it, Google’s launch sits one layer under that surface: the compute and controls a harness runs on; Google says Agent Substrate works with any agent framework or harness, including Claude Code, OpenClaw and Hermes. The framework was published in July 2026 and does not discuss Agent Compute Engine.
What Is Not Established
Google has not published a price or a preview date for Agent Compute Engine. The density and benchmark figures are Google’s own, describe Agent Substrate rather than Agent Compute Engine, and compare against what Google calls traditional compute; we have not seen independent tests. The governance features for laptops and other clouds are described as coming in preview. We did not contact Google. Nothing here is investment advice.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This piece is part of this week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Google announced Agent Compute Engine on 22 October 2026, a service that gives each AI agent its own isolated microVM and already runs Gemini’s agents, with a customer preview due in the coming months. The density figure Google gives, as much as 10x higher compute density than traditional compute, is for the open-source Agent Substrate runtime presented alongside it.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece is based on Google Cloud’s launch post and technical deep dive of 22 October 2026, its 15 September 2026 post on Agent Substrate, Google Cloud’s post on X and our own earlier pieces. The figures are Google’s own. Nothing here is investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
