On October 8, at the Gemini at Work 2026 event, Google Cloud CEO Thomas Kurian officially launched Gemini Agent, positioning it as a "general-purpose work agent" for work scenarios. This is Google's most ambitious product launch to date in the enterprise AI field—AI no longer merely answers questions, but is a "digital colleague" with an independent identity that can run long-term, plan autonomously, and deliver results.
Core Principle: Give It Objectives, Not Instructions
The most fundamental design philosophy of Gemini Agent is summed up in a single sentence: "You give it objectives, not instructions."
Google Cloud CEO Thomas Kurian wrote in an official blog post: "You delegate an outcome, and you come back to the completed work. For an agent to do that, it must be connected to your personal workflows, systems of record, and enterprise controls."
This means users no longer need to tell the AI step by step what to do first and what to do next. They only need to state the final desired result and leave the rest for it to arrange on its own. It can research information, write emails, make PPTs, analyze data, write and run code, and call tools across applications, planning tasks on its own and, when necessary, convening multiple sub-agents to work together.
Four Types of Memory, Like "Onboarding Training" for a New Employee
Gemini Agent has been given a complete memory mechanism comprising four dimensions:
Session Memory: Saves the context of the current task. Even if a task lasts for several days, it can remember which steps have been completed and what work remains.
Semantic Memory: Accumulates knowledge from documents, communications, and collaboration with other agents, organizing it into a structured knowledge base—remembering what products the company has, who on the team is responsible for what, and the specific internal company meaning of a given business term.
Procedural Memory: Records how a job should be done—what data a certain type of report requires, what process to analyze it by, and in what format it should ultimately be delivered. Gemini Agent can even write its own Skills, saving methods learned during work for reuse in subsequent tasks.
Episodic Memory: Records what work has been completed in the past and the corresponding execution experiences.
Google used an analogy: "Gemini will, like a new employee, gradually become familiar with you, your tools, and your team before formally starting work."
Independent Identity: An AI Colleague with an Email Address and Calendar
The most striking design of Gemini Agent is that it can have its own independent Google Workspace account, enterprise email (such as @agents.company.com), calendar, and Drive storage space, and can even appear in the company directory.
Team members can treat it like an ordinary colleague, inviting it to join Google Chat group chats or @-mentioning it in documents. It performs operations under its own identity and leaves its own records—meaning the audit trail shows what "Gemini Agent did," rather than what a certain employee did.
Google has also introduced the Coworker Agent concept for Gemini Agent: enterprises can create a long-lived identity with clear job responsibilities, letting it participate in daily collaboration like any other member of the team.
Multi-Model Routing: It Can Even Call Claude
The underlying model of Gemini Agent is a choice independent of the Agent itself. The system automatically selects the most suitable model for each task.
Notably, at this stage it supports both the Gemini series and Anthropic's Claude series models, and will support other proprietary and open-source models in the future. This open multi-model strategy is not common in Google's own products, reflecting Google's pragmatic stance of "capability first" in the enterprise agent market.
Enterprise-Grade Controls: Permissions, Auditing, and Spending Caps
Gemini Agent was designed from the outset for enterprise governance needs:
- Permission management: Each Agent receives a cryptographically verified identity and follows the principle of least privilege, managing access rights like an employee
- Audit trail: All operations are recorded in audit logs under the Agent's name
- Secure sandbox: Each Agent runs in an independent Agent Sandbox with its own network boundary; all inbound and outbound traffic must pass through the Agent Gateway (AI network firewall)
- Real-time spending caps: Supports intelligent routing and real-time spending limits, letting enterprises flexibly control AI costs
Ecosystem Integration: From Workspace to Third-Party Systems
Gemini Agent is already deeply integrated with the full Google Workspace suite—Gmail, Drive, Docs, Sheets, Slides, Chat, and Calendar—maintaining the same set of memory, skills, and governance controls when used in any channel.
At the same time, it supports connections to a large number of third-party systems including Microsoft 365, Slack, Jira, Confluence, Git, Salesforce, BigQuery, Databricks, Snowflake, and PostgreSQL, and supports secure connections to enterprise internal and external networks via the Model Context Protocol (MCP).
Users can view Gemini's task status through the Task Inbox; the interface displays the reasoning process, tasks assigned to sub-agents, skills and code invoked, and progress updates.
Enterprise Adoption: GSK Shortens a Month of Work to a Few Minutes
Google Cloud disclosed Gemini Agent's early enterprise customers, including On, Shopify, and PayPal. Past customer results for Gemini Enterprise also demonstrated real value: GSK shortened clinical research work that previously took a month to a few minutes; Hotai Motor built its own AI assistant through the no-code Agent Designer, reducing internal data organization time by 30%.
Google CEO Sundar Pichai disclosed a set of key figures at the event: Gemini's monthly active users have exceeded 1 billion, nearly 90% of Fortune 100 companies use Gemini Enterprise, and nearly 500 enterprises each processed more than 1 trillion tokens over the past year.
The launch of Gemini Agent marks the enterprise AI competition entering a new stage—shifting from "whose model is smarter" to "whose AI can work more like a human." Independent identity, four types of memory, multi-model routing, and enterprise-grade governance all point to one goal: turning AI from "a tool that is invoked" into "a colleague with a desk."
Worth noting is Gemini Agent's open support for Claude. When Google is willing to invoke a competitor's model within its own agent framework, it shows that the deciding factor in the enterprise agent market is not the model itself, but the framework, governance, and ecosystem integration capabilities. Whoever can truly embed AI into enterprise workflows, leave auditable records, and respect permission boundaries will win enterprise customers' long-term trust.
This race for the "AI colleague" has only just begun.