Google announces a persistent Gemini agent for work: design permissions before delegating tasks

What Google announced

On October 8, 2026, Google announced a single Gemini agent for work at its Gemini at Work event, in a keynote led by Google Cloud CEO Thomas Kurian. Google describes the agent as a persistent coworker. It runs in the cloud, keeps one set of context and memory across devices and channels, and continues work after you close your laptop. Emma Roth of The Verge independently reported the announcement the same day.

Google attributes a specific set of capabilities to the agent, and each of them remains a vendor statement. The agent can work across Workspace apps such as Gmail, Drive, Docs, Sheets, and Calendar, and from mobile, desktop, and web surfaces, including third-party apps such as Slack and Microsoft 365. Because it runs in the cloud, it keeps the same context across connected devices. Google says it can create temporary, job-specific sub-agents with their own identities for multi-step tasks, and it can act as a coworker agent with a persistent role and a dedicated @agents.company.com email address. Google also describes connections to business systems, fine-grained role-based access approved by an organization's security administrators, agent-attributed audit logs, and real-time spend caps.

Access is limited. The Verge reports that the agent is currently available only to enterprise customers in private preview, and Google's announcement does not state a general availability date. Detailed In Design did not test the product, enroll in the preview, or verify any performance, security, or cost figure. Customer examples in Google's keynote are the vendor's own claims and were not independently verified.

What this means for your business

The section above is reporting. What follows is analysis from Detailed In Design. The announcement matters less for what the agent can do on day one than for the questions it makes practical. A persistent agent that carries context across your systems will eventually touch customer records, documents, and pricing. Before delegating work to any such system, a business owner should be able to answer three questions: which records the agent may read, which actions it may take on its own, and which actions always require a human approval.

Those questions appear in ordinary work. Customer requests arrive by email and phone, job details sit in calendars and spreadsheets, and staff assemble quotes and follow-ups by hand. Someone re-types the details. Someone else checks them. Unrestricted automation could change or disclose records without approval, which is why the discipline matters more than the tool. Software can prepare the work; people still decide.

Detailed In Design could build that discipline as a custom system for a service company: a reviewed request-to-quote workspace. The workspace ingests authorized intake email, extracts the job details, finds the permitted customer documents, and assembles a cited job packet with a draft quote and a draft follow-up. Missing information is routed to office staff. A human approves every price and every outbound message before anything is sent, and the system logs the source records, the reviewer, and the resulting changes. Deterministic software handles the record updates and the authorization checks. This is custom work, separate from Google's product. A Gemini-specific connector would depend on actual preview access, documented APIs, and your tenancy, and the foundational workflow can be scoped without promising Gemini availability. Custom AI, integrations, and workflow automation is the Detailed In Design service this belongs to.

The expected benefit is less retyping and fewer manual handoffs, with clearer responsibility for incomplete work and reviewed customer communications. We make no numerical savings, speed, or accuracy promise. The limitations are real. Cloud execution is not local or offline processing. Drafts may be wrong or omit context. Sensitive pricing, payments, scheduling commitments, and outbound communication retain human approval. The workflow depends on clean records, correct document permissions, review staffing, evaluation against real tasks, audit retention, and cost ceilings. No vendor announcement removes those requirements.

A practical next step

If your team is evaluating agents for work, start with one workflow instead of a platform decision. Map a single process from request to quote, and mark each step where a person must approve. That map becomes the specification for any agent or automation you buy or build, and it is the same map we use to scope custom software around human review gates. Talk to Detailed In Design about a scoped workflow review.

Attribution and scope

Agent capabilities, governance features, and preview status are attributed to Google's announcement and to The Verge's independent reporting. Detailed In Design did not test the product or enroll in the preview. The reviewed request-to-quote workspace is a proposed custom build, not an existing product, a Google integration, or a partnership. This is an attributed synthesis; Detailed In Design conducted no original field reporting.