At the Gemini at Work 2026 conference, Google showed where it wants to take its artificial intelligence. Gemini is evolving from a chat assistant into an agent that plans its own work, connects to corporate systems, and delivers a finished result. It answers questions, handles office tasks, creates images and other media, and writes and runs code. Everything is controlled through a single prompt window.
The new feature is not yet aimed at general users. Google will first make it available to businesses, with other users gaining access later. Google CEO Sundar Pichai noted that more than a billion people use Gemini every month and nearly 90 percent of Fortune 100 companies use Gemini Enterprise. According to him, starting with businesses will help Google address the most difficult challenges posed by such capable agents: security, large-scale operation, and performance. Meta has also introduced its own agents with Project Muse, while OpenAI has launched its Dots service for ChatGPT.
A goal instead of instructions
Google Cloud CEO Thomas Kurian explained the main idea. People no longer need to tell the agent step by step what to do; they only need to give it a goal. Gemini then plans the process itself, selects the appropriate tools, and delivers the completed work directly to documents, inboxes, or development environments.
The agent runs in the cloud, so it remembers the same things on every device. A task that takes hours or days continues even after the laptop is closed. It divides more complex assignments among smaller temporary assistants and manages their work. Meanwhile, people can see how Gemini is reasoning and how far it has progressed in the task overview.
Gemini automatically selects the model on which the agent runs based on the complexity of the task, avoiding unnecessary costs for simple work. Users can also choose the model manually. In addition to Gemini models, Google is offering Anthropic's Claude models from the outset and plans to add open models as well as other proprietary models from different companies later.
Gemini can connect to Google Workspace, Microsoft 365, Teams, Slack, Jira, Git, Salesforce, and ServiceNow, as well as databases such as BigQuery, Databricks, and Snowflake. It works with files on a computer and with any server built on the MCP protocol. People can teach it their own workflows, known as skills, and share them across the company. The agent remembers who it is working with and what it has already done. Google compares this to onboarding a new employee, who first familiarizes themselves with the tools, people, and practices before getting to work.
A colleague with their own email address
Gemini can act as a fully fledged team member. Simply describe its role, and the agent receives its own Workspace account with an email address, calendar, storage, and an entry in the company directory. It knows who belongs to which team and who must approve what. Colleagues can mention it in a chat, send it an email, or share a document with it. When someone mentions it in a comment, it suggests an edit directly in the text and appears under its own name in the version history. It only sees what people share with it.
As a personal assistant, it can, for example, arrange a meeting for the usual team in Gmail, Docs, or Calendar without anyone providing a single name. It identifies the participants from previous communications, checks their calendars, and schedules a time. Gemini can also recognize on its own which work it can take over and offer to do it. When a manager sends an email requesting a presentation on the project's status, Gemini offers to prepare it with a single click.
Monitoring agents and spending
Each agent receives its own verified identity and only the permissions it actually needs. It works in an isolated environment, and all of its activity is monitored by the Agent Gateway network firewall, where a company can configure rules once for all agents. Every step is logged under the agent's name, not a person's. Companies can also set a fixed spending limit for each project. Once the agent reaches it, it stops and can continue only after approval in the administration console.
Where and for whom it will be available
The Gemini agent will run on the web, on iOS and Android phones, on Windows and Mac computers, and from the command line. It will also be available directly in Google Workspace, Microsoft 365, ServiceNow, and Slack. For now, it is a service for Gemini Enterprise business customers. General users will gain access later, but Google has not provided a date.
Early testers included sportswear brand On, Shopify, and PayPal, which sends roughly ten million requests each week across multiple models simultaneously. Gemini Enterprise is used by organizations including BNP Paribas, Merck, and Brazilian bank Bradesco, where reviewing a single document previously took an hour but now takes just five minutes.
Source: techcrunch.com
