Google Cloud has introduced a new Gemini agent for work at its October 8 Gemini at Work event. The idea is broader than asking a chatbot to summarise a document: Google says the agent can plan a task, select tools, draw on company information and return a finished result across the systems people already use.
There is an important boundary to this announcement. Google is describing an enterprise product and its management controls, not declaring that every consumer Gemini user can immediately hand over a multi-application workday to the same agent.
Tools, skills and company context
Google's announcement breaks the agent's capabilities into three overlapping pieces. Tools are connections to business applications and services; skills are reusable instructions for carrying out familiar processes; and context helps the agent understand which documents, conversations and systems matter to a particular assignment.
That combination could be useful for tasks that require more than generating text. A project update, for example, might involve finding records, checking dates and producing a document in an existing collaboration tool. An assistant that can only respond in a chat window still leaves many of those steps to the person asking.
Google says Gemini will operate inside Workspace applications including Gmail, Drive, Docs, Sheets, Slides, Chat and Calendar, and can connect to external business software. The company also describes a tools registry that lets organisations publish approved connections and share skills within teams.
Not every integration implies unrestricted access. Those connectors are supposed to operate within the permissions and policy controls established by the organisation. Exactly which systems work in a given deployment will depend on configuration, rollout and access rights.
One agent, different underlying models
Another part of the pitch is model selection. Google says Gemini can route work among different models depending on the task and its cost, rather than treating a single model as the correct answer to every request.
This separation between the agent and the model underneath it matters operationally. A company may want a sophisticated model for difficult reasoning, but not for every routine formatting or retrieval task. Google describes Smart Routing and spending caps as tools for controlling those trade-offs.
These are product capabilities Google is claiming, not independently established savings for every customer. Real costs will depend on usage patterns, configuration and what the connected systems are asked to do.
Governance is part of the feature set
Google also emphasises identities, authorisation, audit trails, sandboxing and gateways that can constrain how agents access tools. Those details may be less glamorous than a polished demonstration, but they address the main risk of an agent acting on behalf of a worker: an incorrect action can have consequences beyond a bad paragraph of text.
The Verge's reporting places the product within Google's push for a single workplace agent. It notes that the agent's initial availability is a private preview for enterprise users, rather than an instant rollout to every account. That availability caveat belongs beside Google's broad product vision.
The company also describes coworker agents with defined team roles and their own identities. That is an architectural proposal for work that lasts across sessions, not evidence that an organisation can safely remove review and responsibility from automated workflows today.
What changes for businesses now
For organisations already invested in Workspace and Google Cloud, the announcement signals where Google wants to place the next layer of automation: above individual apps, with tools and shared context underneath.
The hard questions are practical. Can the agent see only what it should? Does it retain a clear record of its actions? Can staff stop it or correct mistakes? Does choosing among models keep the cost predictable? Google has described mechanisms meant to answer those questions; implementation and acceptance testing will determine how well they work.
The short version is that Gemini is moving from answering workplace questions toward handling workplace tasks. The news is the announced enterprise agent and its controls. Broad availability, effectiveness and savings still require evidence beyond the launch presentation.
Reporting notes