On 8 October, at the Gemini at Work 2026 conference, Google Cloud announced Gemini agent, a single AI agent for work. You give it a goal rather than step-by-step instructions: it breaks the task into steps on its own, picks the tools it needs, connects to the company’s work systems and hands back a finished result. Corporate customers get the agent first, everyday users later. Google has named neither a price nor a general launch date. According to Decrypt, versions for finance and legal teams are already in preview, with versions for the public sector, healthcare and retail promised later.
Google Cloud CEO Thomas Kurian put the idea like this: work now starts in the prompt box. It sounds like a slogan, but there is a concrete shift behind it. A chatbot answers questions, whereas an agent, as Google sees it, carries out a multi-step task itself and keeps working in the background while you are busy with something else.
What Gemini agent can do
You can hand over a task together with attachments: files, folders or whole “projects”, that is, sets of files and skills for a specific task. From there the agent plans, delegates separate stages to subagents when needed, loads specialised skills and writes code. You can follow all of this in a separate tasks section, which shows how the agent is reasoning, what it has handed to whom and how much is already done.
It runs in the cloud. According to Decrypt, you can close your laptop and the task will keep going: a single job can stretch over hours or even days. The agent puts the result where people already work: in documents, email and development environments. We looked at how agents like this work under the hood (a cloud computer, memory, subagents), based on a Stanford lecture, in our article “How an autonomous AI agent works”.
The list of integrations is long. Inside Google there are Gmail, Docs, Sheets, Slides, Drive, Chat and Calendar. Outside, Microsoft 365, Slack, ServiceNow, Jira, Confluence, Git, BigQuery, Databricks, Postgres and Snowflake. Google also singles out MCP (Model Context Protocol), the open standard for connecting agents to external services: the agent can work with any MCP server, inside the corporate network or outside it. You can launch it from iOS and Android, from a Windows or Mac computer, from the command line and from inside work apps.
An agent employee with its own inbox
The most unusual part of the announcement is agent “colleagues”. A manager describes a role, an events coordinator for example, and Gemini creates an agent for it. The agent gets its own Workspace account: an email address, a calendar, space on Drive and an entry in the company directory. It takes into account who is on which team, what time zones people are in, who signs off on decisions and when people are free, but only within the access it has been given: it sees only what has been shared with it.
You can reach it the same way as a human employee: mention it in a document, email it, share a file with it, add it to a group chat. Every action goes into the audit log and is attributed to the agent, not to the person who called it. For companies this may matter even more than the capabilities themselves: you can always see who did what, an employee or the machine.
Google builds security in several layers. Each agent has its own cryptographically verified digital identity, it runs in an isolated environment called Agent Sandbox, and all its network traffic goes through Agent Gateway, which Google calls a network firewall for AI.
Choice of models: Gemini and Claude
By default the agent picks the model for the task itself, weighing quality against price. But you can also choose the model by hand, and not only from the Gemini family: Anthropic’s Claude models are the first third-party models on the list. Google promises to add open and other proprietary models later.
For Google this is a notable step. A company that promotes its own line-up, from the new Gemini 4 Argon to the lightweight Gemma models, is letting a direct competitor’s models into its flagship work product. The agent’s context, skills and data are kept when you switch models.
Money is tied to that choice too. Agents that work for days and launch subagents can easily burn through a budget, so Google has added cost controls: smart routing between models and real-time spending limits. When a project hits its limit, the agent pauses until a human clicks “continue”.
Why business comes first
Sundar Pichai explained the launch order this way: they start with business in order to solve the hardest questions first, namely security, scale and performance. The logic is clear. An agent with an inbox, access to the CRM and the right to send emails on the company’s behalf is no longer a toy, and its mistakes cost more than a chatbot’s.
There is a commercial calculation as well. According to Pichai, Gemini has more than a billion users a month, and almost 90% of Fortune 100 companies use Gemini Enterprise. Google is betting on this audience: among the agent’s first testers the company names Shopify, PayPal and the Swiss sportswear brand On.
The agent race
Google’s announcement is part of a wider race. Chatbots are learning not only to answer but to act: write code, schedule meetings, book tickets and tables. OpenAI recently showed Dots in ChatGPT, Meta has Muse, and there are now agents that live right inside messaging apps. Where Google differs is that its agent plugs into the email, documents and calendars companies already use every day.
For small teams the main question for now is access terms and price, and until they are announced it is hard to judge the practical value. The direction, though, is clear: if agents like this catch on, the bigger questions will be which work to hand them and how to check the result.
You can try a similar setup without a corporate budget using open agents, for example Hermes Agent.
