Gemini becomes a universal agent, and does the work for you

Written by Jason Miller

Google Cloud has turned Gemini into an enforcer. On October 8th, on the stage of Gemini at Work 2026Thomas Kurian presented Gemini agent: A single agent who accepts “goals, not just instructions,” plans the necessary steps himself, and delivers the finished result. Businesses receive it first; consumers will have to wait.

The distance from the chatbot is measured on three fronts. Gemini agent builds the work plan, loads the necessary skills and generates sub-agents with their own identity to break down complex tasks. The execution remains active in the cloud for days, supported by four types of memory: session, semantic, procedural and episodic.

The choice of model occurs automatically, but the user can intervene and impose the engine he prefers, including Claude from Anthropic. Google promises to expand the selector to open source and proprietary models. On the cost front they arrive smart routingmulti-model orchestration and real-time spending caps in the Cloud Billing Console, with per-project tracking and immediate suspension when the limit is reached.

The AI ​​becomes a colleague with its email, until it overthrows you

The most shocking detail concerns identity. The agent has a Workspace account of your ownincluding email address, and knows the organization chart: who is on which team, the time zones, who has to approve what, other people’s calendars. You summon it with a tag, an email, a share or by inserting it into a group chat, and every action ends up in an audit log attributed to the agent rather than to a person.

The operational perimeter passes through the connectors. Gemini agent communicates with Google Workspace, Microsoft 365, Slack, Jira, Confluence, Git, Salesforce, ServiceNow and databases (BigQuery, Databricks, Postgres, Snowflake), as well as any server MCP inside or outside the company network. Files, folders or entire projects that combine documents and skills can be attached to requests, as the official announcement explains.

Human control is exercised from one tasks inbox which exposes the agent’s reasoning, delegations to sub-agents, loaded skills, product code and progress. Access channels cover iOS, Android, Windows, Mac, command line, Workspace, Microsoft 365 and Slack, with the ability to run it in headlesswithout dedicated interface.

The numbers cited by Sundar Pichai explain the choice to start from the enterprise world: over a billion monthly users for Gemini and almost the 90% of the Fortune 100 already on Gemini Enterprise. Initial testers include On, Shopify and PayPal, while the client list includes BNP Paribas, Bradesco, Merck, Orange Spain, Santee Cooper, SOMPO, Ulta Beauty and Wesfarmers.

Pichai admitted that the postponement of the consumer launch is to solve the “hardest problems around security, scale and performance”. A hefty admission, because the company is handing credentials, company mail, and direct access to production databases to software. Agent Sandbox And Agent Gateway they serve to contain the damage: companies will have to understand how much that safety net is worth before delegating outcomes rather than individual tasks.

Jason Miller

I'm Jason Miller, and I've been passionate about technology and storytelling for over a decade. As a lead writer at Herald Editorials, I strive to bring clarity and creativity to complex tech topics. When I'm not writing, you'll find me exploring the latest gadgets or hiking in the great outdoors.