← All articles

Mikhail Polyakov

Google gave an AI agent a work email address and a place in the company directory: what was announced and what to settle before granting access

Level: BeginnerReading time: 8 min

Topics: AI, Automation, Small business

An office desk with a closed laptop and a mug; in place of the chair stands a translucent glowing figure made of lines and nodes, with glowing threads running from it to the laptop and papers

Short answer: on 8 October 2026, at its own conference, Google Cloud announced the Gemini agent, a single universal agent for work. It lives in the cloud and keeps context regardless of the device, works inside Gmail, Drive, Docs, Sheets and Calendar, is available from the command line, Microsoft 365 and Slack, can create temporary sub-agents for individual tasks, and can become a "coworker agent": with its own account, an email address of the form @agents.company.com, a calendar and a place in the company directory. Two important limits. Everything described is a vendor's account of its own product, and according to The Verge, on the day of the announcement the agent was available only to enterprise customers in private preview. There is no price and no general availability date in Google's post. Below: what exactly was announced, how the agent's memory is arranged, how customer cases differ from verifiable facts, and which questions are worth settling before granting access.

What was announced

The post is based on a talk by Thomas Kurian, the head of Google Cloud. He puts the main announcement this way: the Gemini agent is a single, universal agent for work that answers questions, handles knowledge work, creates images and media, and writes and runs code.

The idea is to replace a set of separate assistants with one. Google's formula: you give it objectives, not instructions, you delegate an outcome and come back to finished work.

How access works: the agent is available from the web, from iOS and Android phones, from Windows and Mac computers, from the command line, from Google Workspace, from Microsoft 365 and from Slack. It can also work without an interface of its own, that is, be built into other applications.

It runs in the cloud. Hence what Google calls continuity: a single set of memory and context regardless of the device, while work that runs for hours or days continues after you have closed your laptop.

Sub-agents and the coworker agent

Two notions from the announcement worth telling apart.

Sub-agents. The agent can assemble, on the fly, a roster of temporary helpers for a particular job, each with its own identity. It coordinates steps with them, including steps running in parallel and in sequence, and such work, by Google's description, can last hours and days.

Coworker agent. This is different: a persistent role that lives longer than one task. Such an agent has its own account, its own email address of the form @agents.company.com, its own persistent storage, and access only to the context that you or your colleagues have given it.

In Workspace it looks like this: the role gets an account of its own with mail, a calendar, Drive and a presence in the company directory. Colleagues work with it in the usual way: they add it to a chat or mention it with an at sign. It acts under its own identity rather than on your behalf, and in a document's version history its edits are signed with its name.

What kind of memory it has

Google lists four kinds of memory, and this is more useful than any promises, because it shows what exactly the agent accumulates:

  • session memory: about the task at hand, even if it runs for several days;
  • semantic memory: a knowledge base it builds itself as it reads documents, talks to people and works with other agents;
  • procedural memory: how a job gets done, including skills it writes for itself;
  • episodic memory: everything it has done before.

Connections to work systems are named separately: Confluence, Microsoft Office, Teams, Slack, Workspace, Git, Jira, Salesforce, ServiceNow, the databases BigQuery, Databricks, Postgres and Snowflake, and files on your own desktop. Plus any Model Context Protocol server, that is, a common way of connecting external sources and actions to AI tools, inside the company network and beyond it.

One more detail that is easy to miss: the model under the agent is a separate choice. Google writes that it picks the model for the job and today uses its own Gemini family and Claude models from Anthropic for that, with other closed and open models promised in the future.

What here is a verifiable fact and what is a vendor's account

All the description of capabilities above comes from Google's post about its own product. How it works cannot be checked from the text.

The same goes for the numbers. Google gives scale: over the last year nearly 500 Google Cloud customers each processed more than one trillion tokens, nearly 80% of all its customers use its AI products, and nearly 90% of the Fortune 100 companies use Gemini Enterprise. How these were counted is not disclosed in the post.

Customer cases are worth reading the same way: BNP Paribas is deploying Gemini Enterprise in an internal assistant for more than 65,000 employees; Bradesco, by its own account, cut document review from an hour to 5 minutes; Orange Spain has deployed more than 1,000 agents of its own; SOMPO has built more than 10,000 agents across 34,000 employees; at the Bunnings chain an internal agent, according to the company, saved staff half a million hours of administrative work. These are statements by companies, published by the vendor. How the savings were counted is not written there.

What is known about availability: according to The Verge, on the day of the announcement the agent was available only to enterprise customers in private preview. There is no price in Google's post, and no date for when it becomes available to everyone.

What to settle before the agent gets access

If you decide to try a tool like this (this one or a similar one), it is useful to answer five questions in advance. They are not about technology but about order inside the company.

  1. What exactly it sees. A coworker agent's access is limited to what has been shared with it. So the scope of access is your decision, and it is better taken before rather than after.
    • Who does it: whoever manages access in your mail and documents.
    • How to check: before the first real task, reconcile the list of folders, chats and systems against what is actually connected, and record the date. If there is no way to check the actual permissions, do not give the agent any work-related data. After that, repeat the reconciliation on a schedule, once a month.
  2. Under whose name it acts. Google writes that a coworker agent acts under its own identity and is signed with its own name in a document's version history. About the other traces of its work, for example in mail and in connections to other systems, the post says nothing, and that is worth finding out separately.
    • Who does it: whoever sets up the agent's account.
    • How to check: open a document the agent edited and find its name in the change history. Separately, ask the vendor in writing which of the agent's actions can be traced and where: in mail, in the calendar, in each connected system. For actions that matter to you and cannot be traced, the conclusion is simple: the corresponding permissions are not granted to the agent.
  3. What it sends outside. The agent can start correspondence, including with external participants: Google gives such an example itself.
    • Who does it: the head of the department where the agent works.
    • How to check: the simplest safe option is to forbid it from corresponding outside on its own. If external letters are needed after all, set up an arrangement where a person reads the recipients and the text before sending, and try it on several letters before putting it into everyday use.
  4. What it accumulates. Google writes that the agent builds a knowledge base itself as it reads documents, talks to people and works with other agents, and that a coworker agent has its own persistent storage. So information about your work is accumulating somewhere.
    • Who does it: whoever is responsible for data at your end.
    • How to check: ask the vendor in writing where this is stored, how it is deleted and what will remain if you stop using the service tomorrow. Keep the answer.
  5. What it costs. Google separately names cost controls: choosing between models, smart routing and real-time spend caps. There is no price and no payment scheme in the post, so the counting has to be done on your own tasks.
    • Who does it: whoever pays.
    • How to check: set a spend cap before the first real task, and look at the first month's bill next to the number of tasks the agent closed.

I wrote separately about a similar step by Microsoft, which gives Copilot access to local files: Copilot and access to files in Windows. And on what happens when an agent works around the rules it was given, there is another piece: When an AI agent works around the rules.

What not to expect from this news

  • That you can buy it today. According to The Verge, on the day of the announcement the agent was open to enterprise customers in private preview.
  • That the price is known. There is no price in Google's post.
  • That the result will repeat at your end. The hours and minutes saved in the cases are the companies' own words about their own processes, without disclosing how they were counted.
  • That the agent replaces an employee. The announcement describes a role that is given access and tasks. Who is responsible for the results of its work is up to you.

In summary

On 8 October 2026 Google Cloud announced the Gemini agent: a single agent for work living in the cloud, accessible from Workspace, Microsoft 365, Slack and the command line, with four kinds of memory, able to assemble temporary sub-agents and to work as a coworker agent with its own mail and a place in the company directory. It picks the model for the job itself, today from the Gemini family and Claude models. All of this is the vendor's description: according to The Verge, on the day of the announcement access was limited to a private preview, and there is no price in the post. For a company the main question here is not about capabilities but about order: what the agent sees, under whose name it acts, what it sends outside, what it accumulates and what it costs.

I work on AI agents and automation. If you would like to see my projects or discuss your own task, take a look at my portfolio.

Sources