Google Cloud announced the Gemini agent on October 8, 2026. In its keynote account, the company describes a work agent that can use business tools, retain context across apps and continue longer tasks in the cloud. Google also outlined per-agent identities, network policy controls and project-level spending caps. These are company descriptions, not independent evidence that every control works in a particular deployment.
How the agent is designed to work
Google says the agent can take an objective, call reusable skills and tools, and coordinate temporary sub-agents for multi-step work. It also describes model selection across its Gemini models and Anthropic's Claude models. The announcement does not show how reliably those parts work together in a customer's environment.
The main operational question is narrower than whether an agent can complete a task: which systems can it reach, whose authority does it use, and how can administrators inspect its actions?
Identity and network boundaries
Google says each agent has its own identity and least-privilege permissions. It says actions are logged under that identity, including activity in virtual machines the agent creates to run code. For external systems, Google describes mapping and propagating the agent's identity through standards such as OAuth.
The company also describes an Agent Sandbox with a network boundary and an Agent Gateway intended to apply policy to traffic into, out of and between agents. A buyer should check the permissions actually granted to an agent, try an action that policy should deny, inspect the resulting logs and establish who can stop a running task. Those are evaluation steps, not test results supplied by Google.
A project cap is different from a token tier
Google says administrators can set a hard limit on a project's AI spending in the Cloud Billing Console. According to the company, the cap tracks token use and sandbox costs, pauses that project's agent when triggered, and allows an administrator to resume it. Google also says Smart Routing selects a model for each workload. Neither the announcement nor those control descriptions establish an independently measured cost saving for a reader's workload.
Teams should test a representative task against the cap, including model choice and sandbox use, before using Google's efficiency claims in a budget forecast.
Rollout details still matter
The October 8 keynote post does not give a general rollout date or a capability-by-capability availability table for the Gemini agent. Google separately labels its financial-services and legal specializations as previews. Organisations evaluating the broader agent should confirm access to the specific connectors, identity controls, gateway policies and billing limits they need. The practical value depends on those controls working together in the deployment a team can actually use.
