Yesterday at its Gemini at Work 2026 event, Google announced a new Gemini agent for work: one agent meant to handle many kinds of tasks across the tools a business already uses. It is in limited preview and aimed at larger organisations first, so most smaller teams cannot try it yet. But the announcement is a useful look at where workplace AI is going, and at the questions any team should ask before letting an agent into their systems.
What Google announced
Google describes the Gemini agent as a universal agent for work. You give it an objective rather than a list of steps, and it plans the work, uses tools and delivers a finished result such as a document, a spreadsheet, a summary or a piece of code. It combines what used to be separate things: chat, autonomous task work and coding.
Reports say it runs in Google’s cloud, so it keeps context across web, mobile and desktop. For bigger jobs it can create temporary helper agents, each tagged with its own identity so their actions can be audited. It can also act as a standing team member with its own email address, calendar and file storage.
Which tools and models it works with
This is the part that stands out. At launch, the agent is reported to work across Google Workspace, Microsoft 365 and Slack, and to connect to tools such as Jira, Salesforce and ServiceNow, with support for the Model Context Protocol, a common way of plugging in other software.
It currently uses Google’s own Gemini models and also Anthropic’s Claude models, choosing between them based on the task, or letting you pick. Google reportedly says other models will follow. For a business, that mix matters: it suggests you may not have to bet everything on one vendor’s model.
The controls Google highlights
As with any agent, the safeguards matter as much as the features. Coverage of the launch mentions:
- Each agent has its own identity, with access limited by role.
- Actions are logged under the agent’s name rather than a person’s.
- Agents can reportedly see only the files and folders explicitly shared with them.
- Administrators can write company-wide rules that apply to every agent, such as blocking access to documents marked as restricted.
- Administrators can set a spending cap per project, after which the agent pauses until someone resumes it.
These are exactly the right kinds of control. Whether they work as smoothly in practice as on a stage is something only real use will show.
What is not confirmed
This is early, and several things are unclear. Treat the following with caution.
- Availability. It is reported to be in a private preview limited to a small number of enterprise customers. Google has said broader access is coming, but we have not seen a confirmed date.
- Price. Google has reportedly said it is included in Gemini Enterprise at no additional charge where that product is available, while usage may be metered. Some third-party figures for per-user pricing are circulating, but Google has not confirmed them, so do not budget around them.
- Industry versions. Versions for financial services and legal work are said to be in preview, with others planned.
Details may change between the announcement and a general release.
What it means for smaller teams
Google has framed this mainly as an enterprise product, so the title question has a short answer: for now, smaller teams mostly get a preview of what is coming, and a chance to prepare.
Honestly, you probably cannot use this yet, and you do not need to. Three lessons are worth taking now.
Agents are becoming part of the office suite. The pattern is clear across the industry: agents are moving into the tools you already pay for, such as email, documents and chat. Within a year or two you may well be offered one by your existing provider, rather than having to go looking.
Identity and logging are the new basics. The strongest part of this announcement is the idea that every agent has its own identity, limited access and a readable log. You can ask for the same from any tool you adopt, and you can build the habit now, even with simple tools.
Cost control is part of the product. Spending caps are included for a reason: agents that take many steps can run up usage quickly. Whatever you choose, ask how you can limit and monitor spend.
What to do now
- Find out which Google or Microsoft plan you are on, and whether your provider has said when agent features will reach it.
- Tidy your file sharing. An agent that can see everything shared with it will expose any messy permissions you already have. Fix them first.
- Write a short agent policy: who may enable one, what it may connect to, and which actions need human approval.
- Choose two or three low-risk tasks you would like to hand over in future, and measure how long they take today so you can judge any benefit honestly.
- Ask your current providers, in writing, how agent features will be priced, logged and switched off, so you are not surprised later.
- Wait for real-world results before committing. A preview is a demonstration, not a track record.
The bigger story is not one product but a direction: AI that carries out work across your tools, with tighter controls than the early chatbots had. If you would like help getting your permissions, policies and processes ready, so that you can adopt these tools confidently when they arrive, it is a good use of an hour or two with an experienced adviser.