Google Launches Gemini Agent for Work: How the New Universal AI Agent Works

Gemini Agent is Google Cloud’s new universal AI agent for work, announced on October 8, 2026 at Gemini at Work 2026. Unlike a standard chatbot that mainly responds inside a single conversation, the new agent is designed to accept an objective, plan the work, use tools, connect to business systems, coordinate sub-agents and return finished work inside the documents, inboxes and developer environments employees already use.

Google is positioning the agent as a cross-application work layer rather than a model-specific product. It works inside Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, but it can also operate through Microsoft 365, Slack, command-line tools and third-party applications. It can run for hours or days in the cloud, preserve memory across devices and channels, and choose among different AI models for each task.

That last point is important: Gemini Agent is not the same thing as a single Gemini model. Google says the agent can currently orchestrate work across Google’s Gemini model family and Anthropic’s Claude models, with additional private and open models planned in the future.

Quick take: Gemini Agent is Google’s attempt to become the universal work agent for the enterprise. It combines persistent execution, multi-agent orchestration, cross-app access, coworker agents, model routing, enterprise identity, audit logs, agent sandboxing and real-time spend controls in one system.

What is Google Gemini Agent?

Gemini Agent is a persistent enterprise AI agent that can handle questions, knowledge work, content creation, media generation and coding from one interface.

Google describes the interaction model differently from a normal assistant: you give it an objective rather than a sequence of detailed instructions.

The agent can then:

• plan the work;
• select tools and skills;
• connect to business systems;
• choose an AI model;
• create temporary sub-agents;
• run tasks in parallel or sequence;
• continue working after the user closes their laptop;
• bring the finished result back into Workspace, developer tools or connected applications.

That puts Gemini Agent directly into the same broader category as persistent systems such as OpenAI Dots, Meta Muse and Microsoft’s increasingly agentic Copilot stack.

Gemini Agent works across Google Workspace

Inside Google Workspace, Gemini Agent can work directly in:

• Gmail;
• Drive;
• Docs;
• Slides;
• Sheets;
• Chat;
• Calendar.

Google says the same memory, skills and enterprise controls can carry across those surfaces.

That means the user does not need to re-explain the same project every time they move from email to a spreadsheet or presentation.

For example, Google says an employee can ask Gemini to organize a meeting with the usual regional event team. The agent can identify those people from previous collaboration context, check calendars and start the coordination process without the user manually listing every participant.

In another example, Gemini can research a market, build a financial model in Sheets and then create a presentation from the analysis.

Gemini Agent also works with Microsoft 365 and Slack

The agent is not limited to Google’s own ecosystem.

Google says Gemini can be accessed through:

• Google Workspace;
• Microsoft 365;
• Slack;
• command-line interfaces;
• web;
• iOS and Android;
• Windows and Mac;
• third-party applications;
• headless agent environments without a dedicated interface.

The tools layer goes further. Google lists integrations with systems such as Confluence, Microsoft Office, Teams, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres and Snowflake.

Gemini can also connect to Model Context Protocol servers inside or outside the company network.

That is strategically important because enterprise software is heterogeneous. A company may use Gmail for email, Slack for communication, Microsoft Excel for finance, Salesforce for CRM and Jira for engineering. Google is trying to make Gemini Agent operate above that fragmentation.

How Gemini Agent works

Google describes the architecture around several core ideas.

This is closer to an operating layer for AI work than a traditional chatbot interface.

Gemini Agent can run for hours or days

Persistent execution is one of the defining features.

Because Gemini runs in the cloud, Google says it can keep working even after the user closes the laptop or switches devices.

That makes it suitable for long-running tasks such as:

• multi-stage research;
• data analysis;
• large code-generation workflows;
• financial analysis;
• document review;
• scheduled reporting;
• multi-system business processes.

The user does not need to keep one browser tab alive while the work continues.

This places Gemini in direct competition with the “always-on” agent model covered in Vynula’s OpenAI Dots analysis.

Gemini Agent can create temporary sub-agents

The new system also supports multi-agent orchestration.

Google says Gemini can dynamically create a roster of temporary, job-specific sub-agents for complex tasks.

One sub-agent could research data, another could analyze it, another could build a presentation and another could check the result.

Those steps can run:

• in parallel;
• sequentially;
• for hours;
• or over multiple days.

The parent Gemini Agent coordinates the workflow.

This is a major shift from the simpler pattern where one model receives one prompt and produces one response.

What are Gemini coworker agents?

Gemini coworker agents are persistent AI agents designed to behave more like team members than temporary task workers.

Google says a coworker agent can have:

• its own identity;
• its own Workspace account;
• its own email address;
• its own calendar;
• its own Drive storage;
• presence in the organization’s directory;
• a defined role;
• persistent responsibilities across sessions.

Instead of using the human employee’s identity, the coworker agent acts under its own.

This is an important enterprise design decision because it makes agent actions easier to separate, audit and govern.

Coworker agents get their own company email

Google gives a striking example of how far the identity model goes.

A coworker agent can receive a dedicated email address such as an @agents.company.com identity.

It can then participate in work the way a human teammate might:

• join a Chat space;
• be @mentioned;
• post a completed document;
• reply to comments;
• suggest document edits;
• appear under its own name in version history.

Google says coworker agents only see the context explicitly shared with them or available through permissions granted to their identity.

This is much safer than simply giving an AI agent the full access rights of the employee who created it.

Gemini Agent has four types of memory

Google says the new agent maintains four categories of memory:

• Session memory: what the agent needs for the current task, including work that continues for days.
• Semantic memory: structured knowledge about documents, people and organizational context.
• Procedural memory: information about how tasks are performed, including skills the agent learns or creates.
• Episodic memory: history of what the agent has done before.

The company compares this to onboarding a new employee who gradually learns how the organization works.

This long-term context is a major part of Google’s attempt to make Gemini more useful over time rather than forcing users to repeatedly brief it.

Gemini Agent is not tied to Gemini models

Despite the name, Gemini Agent is not limited to Google’s own models.

Google says the agent can currently route tasks across:

• Google Gemini models;
• Anthropic Claude models.

Other private and open models are planned for the future.

This is one of the most strategically important parts of the launch.

Google is separating the agent from the model.

The agent holds the business context, tools, memory, security controls and workflow. The model becomes a replaceable reasoning engine underneath it.

Why Google is using Anthropic Claude inside Gemini Agent

Different models are strong at different tasks.

Google argues that the best model for a job is not always the biggest or most expensive one.

A simple classification task may use a cheaper, faster model. A difficult coding or reasoning task may benefit from a stronger model.

By routing work dynamically, Google aims to improve both performance and cost efficiency.

Reuters confirms that the current system can use both Gemini models and Anthropic Claude models.

This also reduces lock-in at the workflow layer. If one model improves dramatically in six months, a company may not need to rebuild all of its skills, permissions and connections just to use the new model.

Smart Routing tries to reduce enterprise AI costs

Model routing is also central to Google’s cost strategy.

Google calls one component Smart Routing.

The system can triage workloads and choose the model that provides the best performance at an appropriate cost.

For example:

• a basic summarization task might go to a cheaper fast model;
• a complex financial workflow might go to a frontier reasoning model;
• a coding subtask might be assigned to whichever model performs best for that class of work.

This matters because enterprise agents can consume enormous token volumes when they run many autonomous loops.

Real-time spend caps can pause agents

Google also announced real-time AI spending controls.

Companies can set a hard spending cap for a project in Cloud Billing.

Gemini monitors:

• token usage;
• sandbox costs;
• project-level spending.

If a project reaches the configured limit, the agent can be paused automatically.

The administrator can then decide whether to resume it.

This is a practical response to one of the biggest enterprise-agent risks: an autonomous system can continue making model calls long after a human would have stopped.

Gemini can connect to MCP servers

Google says Gemini Agent can securely connect to Model Context Protocol servers inside or outside the company network.

MCP gives AI systems a standardized way to discover and call tools, resources and services.

Vynula’s Model Context Protocol guide explains how that ecosystem works.

For Gemini Agent, MCP support means organizations do not have to wait for Google to build every possible integration.

They can expose internal systems through compatible servers and allow the agent to use them within enterprise policy.

Tools Registry and Skills Registry

Google is also formalizing how companies distribute agent capabilities.

The Tools Registry allows organizations to publish approved tools that other teams can discover and use.

The Skills Registry does something similar for reusable workflows and instructions.

A skill might define how the company:

• prepares a quarterly report;
• reviews a contract;
• runs an onboarding workflow;
• investigates a customer complaint;
• prepares an investment memo.

Instead of every employee rebuilding the same prompt, the company can publish a reusable skill.

Gemini Agent security: every agent gets an identity

Google’s security model starts with identity.

The company says every agent receives its own cryptographically attested identity and least-privilege permissions.

This allows an administrator to answer:

• Which agent performed this action?
• What permissions did it have?
• Which data could it access?
• Which external system did it connect to?
• What did it do?

The identity is carried into logs and even into virtual machines created to run code on the agent’s behalf.

This is important because enterprise AI cannot be governed properly if every automated action simply appears as though the human employee performed it.

Every Gemini Agent action is auditable

Google says agent actions are written to audit trails and attributed to the agent identity.

Organizations can monitor those logs using Google Cloud observability tools.

This creates a clearer chain of responsibility than systems where agents silently borrow a user’s identity.

The design reflects a broader shift across the industry: AI agents are becoming first-class enterprise identities rather than invisible scripts.

Agent Sandbox isolates execution

Google also says Gemini agents execute tasks inside an Agent Sandbox with its own network boundary.

This is designed to limit the damage that can occur if:

• an agent executes unsafe code;
• a tool behaves unexpectedly;
• a prompt-injection attack changes agent behavior;
• a dependency is malicious;
• the model makes a bad decision.

Sandboxing does not make an agent perfectly safe, but it can reduce the blast radius.

Agent Gateway acts like an AI network firewall

All traffic between agents and external systems can pass through Google’s Agent Gateway.

Google describes it as an AI network firewall.

Companies can define organization-wide policies such as:

“Agents may not open documents classified Need to Know.”

The policy can then be enforced centrally instead of being manually configured for every agent.

This type of centralized control will likely become increasingly important as companies deploy hundreds or thousands of agents.

Gemini Agent vs OpenAI Dots

Google’s new system enters a market that is moving quickly.

OpenAI introduced OpenAI Dots as persistent agents that can keep working across apps and cloud environments.

The competition is increasingly less about which chatbot is smartest and more about which company can build the best agent operating layer.

Gemini Agent vs Meta Muse

Meta Muse is more consumer- and personal-assistant oriented.

Vynula recently covered how Meta Muse is moving into physical devices.

Muse can shop, book travel, send emails and perform other personal tasks.

Gemini Agent is being positioned more directly around enterprise workflows, corporate identity, data governance and organizational systems.

Both systems share the idea of persistent memory and long-running work, but their commercial centers are different.

Gemini Agent vs Microsoft Copilot

Microsoft is also turning Copilot into a more autonomous work layer.

Vynula’s Microsoft Copilot overhaul analysis covers the Home, Code and Autopilot structure.

Microsoft’s strength is its ownership of Office, Windows, Azure, Entra and enterprise management infrastructure.

Google’s response is to make Gemini portable across both Google and third-party work environments.

That means the competition could become:

• Microsoft: deep control of Windows + Office + Azure;
• Google: universal work agent across Workspace + external platforms + multi-model routing;
• OpenAI: broad agent platform around ChatGPT and Dots;
• Meta: personal agent across consumer services and devices.

Gemini Enterprise agent specializations are already emerging

Google is also creating specialized versions of Gemini for specific industries.

Gemini for Financial Services and Gemini for Legal are currently in preview.

Versions for:

• government;
• healthcare;
• retail

are listed as coming soon.

Financial Services includes workflows such as investment research, credit analysis and risk modeling.

Google says the system can draw on data from FactSet, LSEG, S&P Global, SEC filings and internal repositories.

Gemini for Financial Services includes 50+ foundational skills

Google says the finance version includes more than 50 foundational skills.

It is designed to expose:

• confidence scores;
• methodologies;
• data lineage;
• source citations.

Those features are particularly important in regulated work where an answer without an audit trail may be unusable.

Gemini for Legal inherits matter-level permissions

For legal teams, Google says Gemini can inherit existing permissions and ethical walls from document-management systems such as NetDocuments and iManage.

This is critical because a law firm cannot allow an AI agent working on one client matter to casually access confidential files from another.

The legal version is therefore as much about access boundaries as model intelligence.

Google says nearly 80% of Cloud customers use its AI products

Google used the event to emphasize that this is not only an experimental platform.

The company says nearly 80% of Google Cloud customers now use its AI products and nearly 90% of Fortune 100 companies use Gemini Enterprise.

These are Google-reported adoption figures, not independently audited market-share numbers.

Google also says nearly 500 Cloud customers each processed more than one trillion tokens over the past year.

Real customer examples show where Google wants Gemini Agent to go

Google highlighted several enterprise deployments around Gemini Enterprise.

Examples include:

• Orange Spain deploying more than 1,000 custom agents;
• SOMPO building more than 10,000 custom AI agents across 34,000 employees;
• Qatar University deploying more than 2,000 agents;
• Tata Steel deploying more than 300 agents in nine months;
• Bunnings reporting that an internal agent saved staff half a million hours of administrative work.

These are company-reported examples and should not be interpreted as guaranteed results for every organization.

Gemini Agent can work directly with enterprise data

Google is also connecting agents more deeply to structured and unstructured enterprise data.

The company announced capabilities around:

• Knowledge Catalog;
• Smart Storage;
• Borderless Lakehouse;
• BigQuery;
• Databricks;
• Snowflake;
• SAP;
• Salesforce;
• Amazon S3;
• Azure Data Lake.

The goal is to let agents use business data without requiring every organization to copy everything into one new AI database.

The agent can generate and run code

Gemini Agent also includes developer workflows.

Google says it can:

• generate code;
• run code;
• create notebooks;
• generate PySpark;
• debug pipelines;
• work with Git and Jira;
• coordinate technical sub-agents.

This broadens the product beyond office productivity.

Google is positioning one agent layer across knowledge work, analytics and software development.

How Gemini Agent fits with Gemini 4 Argon

The agent layer is separate from the underlying frontier-model roadmap.

Vynula recently covered Gemini 4 Argon, Google’s frontier model for difficult professional work.

Google’s Gemini at Work announcement describes Argon as one of the models available at the infrastructure layer, alongside Flash, Omni and Gemma.

Gemini Agent can select among models according to the task.

That means the long-term value of the agent may depend less on any one model release.

If Google replaces Argon with a better frontier model next year, the same tools, memory, skills and business context can remain attached to the agent layer.

Google AI agent for work: why this could matter more than a new model launch

Model releases get attention because benchmark changes are easy to measure.

But enterprise buyers often care more about whether an AI system can safely complete a real workflow.

That requires:

• identity;
• permissions;
• memory;
• connectors;
• tool execution;
• model routing;
• security;
• auditability;
• cost controls.

Gemini Agent is Google’s attempt to package those elements into one work layer.

What Gemini Agent does not prove yet

The launch is ambitious, but several important questions remain.

• Reliability: Can the agent complete long multi-app workflows consistently?
• Routing quality: Does model selection actually choose the best model often enough?
• Cost: How expensive are multi-day autonomous workflows at enterprise scale?
• Permissions: Can organizations configure least privilege without overwhelming administrators?
• Memory accuracy: Will long-term context remain useful without accumulating outdated assumptions?
• Cross-platform depth: How deeply can Gemini operate inside Microsoft 365, Slack and third-party systems compared with native assistants?
• User trust: Will employees be comfortable working alongside persistent coworker agents?

Those questions matter more than the launch demo once organizations begin using agents on high-value work.

Vynula analysis: Google is trying to own the layer above the model

The most important part of Gemini Agent is not Gmail integration or even coworker identities.

It is Google’s decision to separate the agent platform from the underlying AI model.

This is strategically powerful.

If models become increasingly interchangeable, the winner may not be the company with the best benchmark score on a particular week.

The winner may be the company that owns:

• the user’s work context;
• the organizational memory;
• the skills;
• the tools;
• the permissions;
• the audit trail;
• the identity;
• the routing layer.

Google’s architecture is effectively saying: the model can change; the work layer stays.

That is why allowing Claude inside Gemini Agent is not a weakness. It may actually be the point.

Google is attempting to make itself the operating layer that chooses the model, rather than forcing every company to rebuild its enterprise AI stack around whichever model happens to be strongest at the moment.

In Vynula’s view, this is where the enterprise AI race is heading.

OpenAI Dots, Microsoft Copilot, Meta Muse and Gemini Agent all point in the same direction: models are becoming components inside persistent agent systems.

The competitive advantage increasingly moves upward from model intelligence toward context, execution, identity and control.

If Google can make Gemini work reliably across Workspace, Microsoft 365, Slack, enterprise databases and developer tools, it could become much harder for companies to replace the agent layer than to replace the model underneath it.

That may be the real strategic objective of Gemini Agent.

What to watch next

• Broader availability: how quickly Gemini Agent becomes available across Gemini Enterprise customers and regions.
• Microsoft 365 depth: whether external integrations approach the quality of native Workspace integration.
• Claude routing: when Google expands model choice beyond Gemini and Claude.
• Coworker adoption: whether companies actually create persistent AI employees with their own identities.
• Agent governance: how well Agent Gateway and sandbox controls hold up under real attacks.
• Cost controls: whether Smart Routing materially lowers autonomous-agent spend.
• Industry versions: Government, Healthcare and Retail launches.
• Competitive response: what OpenAI, Microsoft and Anthropic add to their own enterprise agent platforms.

Frequently asked questions

What is Gemini Agent?

Gemini Agent is Google Cloud’s universal AI agent for work. It can answer questions, complete knowledge work, create content, generate and run code, use business tools and work across multiple applications.

Is Gemini Agent the same as a Gemini model?

No. Gemini Agent is an orchestration and execution layer. Google says it can currently use both Gemini models and Anthropic Claude models, with additional models planned.

Which apps can Gemini Agent use?

Google lists Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, along with Microsoft 365, Slack, command-line tools and third-party systems such as Salesforce, ServiceNow, Jira and databases.

What are Gemini coworker agents?

Coworker agents are persistent AI team members with their own organizational identity, email, calendar, Drive storage and role-based access.

Can Gemini Agent run after I close my laptop?

Yes. Google says persistent tasks can continue running in the cloud for hours or days.

Does Gemini Agent support Claude?

Yes. Google says the agent can currently orchestrate tasks across Google Gemini models and Anthropic Claude models.

Does Gemini Agent support MCP?

Yes. Google says it can connect to Model Context Protocol servers inside or outside an organization’s network.

How does Google secure Gemini Agent?

Google says each agent receives its own identity and least-privilege permissions. Actions are logged, execution runs inside Agent Sandbox environments and network traffic can be controlled by Agent Gateway policies.

Which industry-specific Gemini agents are available?

Financial Services and Legal are in preview. Government, Healthcare and Retail are listed as coming soon.

Primary and independent sources

• Google Cloud — Welcome to Gemini at Work 2026: Introducing the Gemini agent
• Google — Google Cloud introduces the Gemini agent
• Reuters — Google Cloud introduces Gemini agent for work as AI race heats up

Reporting note: Product architecture, integrations, security controls, customer adoption and performance claims are attributed to Google unless otherwise stated. Reuters independently confirms the launch, cross-platform support, Claude model routing, coworker-agent identities and industry-specific preview plans. Vynula Analysis represents editorial interpretation. Images are illustrative Unsplash photography, not official Google interface screenshots.