Gemini Agent: What Google’s Work AI Does and Who Gets It

Google wants your next work request to start in one prompt box, even when Anthropic’s Claude does part of the job. Announced on October 8, 2026 at Gemini at Work 2026, Gemini agent is Google Cloud’s work agent that takes your goal, plans the tasks and assigns work to helper agents and models. It is in private beta, with Constellation Research reporting a public release around late October or early November.

You see, the pitch goes beyond a useful reply. You hand over a result you want, and the agent works towards delivering it inside your existing apps. Here are the key facts.

Detail	What we know
Announced	October 8, 2026 at Gemini at Work 2026
Made by	Google Cloud
Status	Private beta, with general availability expected in late October or early November according to Constellation Research
Models	Google Gemini and Anthropic Claude at launch
Where it works	Web, iOS, Android, Windows, Mac, CLI, Workspace, Microsoft 365 and Slack
Price	Usage-based billing, no separate SKU or add-on reported, and no published list price
Industry editions	Financial Services and Legal are in preview

What the Gemini Agent Actually Does

You give Gemini agent an objective, rather than instructions for every step. Google presents it as a single agent with one API and a shared prompt box for several types of work. It plans the task, calls tools and business systems, and delivers output in documents, email or developer tools. Here is what you can ask it to handle.

• Questions and knowledge work. You can request answers and delegate work that uses information held in your connected business systems.
• Content and media. You can ask for written material, images and other media through the same agent.
• Coding. You can delegate both writing code and running it.
• Recurring or triggered tasks. You can assign a job directly, schedule it or have an event start it, according to Constellation’s reporting.
• Longer assignments. Cloud execution lets jobs continue for hours or days after you shut your laptop.

In Google Cloud’s Gemini at Work announcement, the aim is to move work into the prompt window and let you return to a completed assignment.

Sub-Agents, Coworker Agents and Four Kinds of Memory

A helper created for one assignment differs from an agent that keeps a role in your team. Gemini agent supports both, with memory that carries information across devices and long-running jobs. The differences become clearer when you look at identity, access and what each memory keeps.

Temporary helpers for a single job

For a particular assignment, Gemini agent can create temporary sub-agents with separate identities. They can work at the same time or take turns through successive steps, including tasks that run over several days.

Agents that join your team

Coworker agents keep a defined role and their own identity, storage and email address using the @agents.company.com format. In Workspace, that includes a separate account with email, Calendar, Drive and a directory entry.

You can see their own names against edits in version history. Their access is limited to context people share with them, so joining your team does not give an agent unrestricted access to company information.

How it remembers your work

Google describes an agent learning about you, your tools and your team during onboarding, much as a new employee would. Its four memory types serve different purposes.

Memory type	What it stores
Session	Context for the active assignment, including work continuing across days
Semantic	Knowledge gathered from documents, conversations and other agents
Procedural	Methods for completing tasks, including reusable instructions the agent develops
Episodic	A history of the actions the agent has taken

Why a Google Agent Runs Claude Models

You do not have to use a Google model for every task. Gemini agent can send work to Gemini or Anthropic Claude today, with additional private and open models planned later. The reason comes down to how Google selects a model for each job.

Smart Routing aims to match task performance with the lowest cost. Your administrator can leave model selection automatic or restrict expensive options, according to Constellation Research.

Analyst Holger Mueller welcomed this model choice because it lets companies use language models they have already purchased within Gemini agent.

One detail remains unanswered. Google has not identified the Claude versions available through the agent. You can read about Anthropic’s latest Claude Haiku release, but that does not establish which Claude model your work agent will call.

Where You Can Use It and What It Connects To

You can reach Gemini agent through several devices and workplace channels, while its cloud context follows you between them. It can also run without a user interface inside third-party apps. These are the access points and connections named in the research.

Area	Supported options
Devices and interfaces	Browser, iOS, Android, Windows, Mac and command line
Workplace channels	Google Workspace, Microsoft 365 and Slack
Workspace apps	Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar
Business tools	Confluence, Microsoft Office, Teams, Git, Jira, Salesforce and ServiceNow
Data and files	BigQuery, Databricks, Postgres, Snowflake and local desktop files
Other connections	Any MCP server
Reusable skills	Google’s shared library, a company registry and personal instructions

Within Workspace, Google gives you a few concrete examples.

• Arrange a meeting. Ask it to book time with your usual group without entering each person’s details.
• Move between formats. Have it build a financial model in Sheets and then turn that work into a presentation.
• Delegate a suggested task. Workspace Intelligence can offer work you hand over with one click.

Gemini Agent Pricing and Availability

You cannot yet attach a published price to Gemini agent. Its beta status, reported billing structure and release estimate need to be read separately. Here is what the supplied sources establish.

Item	Status	Source
Access now	Private beta	Constellation Research
General release	Expected late October or early November 2026	Constellation Research
Billing	Consumption-based, without a separate SKU or add-on	Constellation Research
Workspace plans	Continue, with agent consumption details still pending	Constellation Research
List price and free tier	Neither announced in the supplied research	Supplied research
Spend limits	Google describes hard project caps, while Constellation reports project and individual caps coming within weeks	Google Cloud and Constellation Research
Industry editions	Financial Services and Legal previews, with Government, Healthcare and Retail to follow	Google Cloud

No extra add-on does not mean free usage. Your bill depends on consumption, and the exact Workspace charging model remains unannounced.

Google says an agent pauses when a project reaches its spending limit and can resume with one click. Before budgeting for it, you still need the published rates and confirmation of which spending controls are available to your organisation.

Gemini Agent vs Gemini Spark vs Agent Platform

Similar names, different jobs. Do you need personal help, a company work agent or tools to build an agent yourself? Here is how Google’s three products compare.

Product	Who it is for	What it does	How you get it
Gemini agent	Businesses and their employees	Handles work objectives, creates helper agents and routes tasks between Gemini and Claude	Google Cloud private beta, with usage-based billing reported
Gemini Spark	People wanting a personal assistant	Runs around the clock and handles tasks through Gmail, Docs, Slides and partner apps	Announced at Google I/O in May 2026, initially for testers and Google AI Ultra beta users in the US
Gemini Enterprise Agent Platform	Developers building business agents	Provides agent development, deployment and governance tools, with more than 200 foundation models	Through Google’s developer platform, including ADK and Agent Studio

Spark connects to partners including Canva, OpenTable and Instacart. You choose its app connections, and it asks before actions such as spending money or sending email.

Google’s consumer plans list AI Plus at $4.99 a month, Pro at $19.99 and Ultra from $99.99. Those prices do not establish Gemini agent’s business rates. You can also read about Google’s Gemini 4 Argon model.

How It Stacks Up Against Copilot, Dots and Muse

You have several competing agents to consider, but their access, models and billing differ. Calling them the big four is an informal comparison, rather than an official category. For your decision, these are the differences that matter.

Agent	Company	Launched	Models	Where it works	Price info
Gemini agent	Google Cloud	Announced October 8, 2026	Gemini and Claude	Apps, web, desktop, CLI, Workspace, Microsoft 365 and Slack	Consumption billing reported, with no list price
Microsoft 365 Copilot agents	Microsoft	Available, with no launch date given in the supplied pricing source	Model choice listed without provider names	Microsoft 365	Copilot Business normally $21 per user monthly, with an $18 first-year promotion and annual commitment
Dots	OpenAI	Announced September 29, 2026	GPT-6 Astra	ChatGPT, Slack, Teams and over 4,000 apps through plugins	One agent included with Pro and Business Premium, with extra-agent rates pending
Muse	Meta	US launch September 8, 2026	The supplied launch research does not identify a model	Personal agent on a dedicated cloud computer	Reported app tiers of Free, Power at $20 monthly and Maximum at $100 monthly

Copilot lists Researcher, Analyst and Facilitator agents, while Dots gets its own cloud computer. Muse can shop, arrange travel, email and make payments. For context, see Meta’s new Muse Wonder model and ChatGPT’s recent interface changes, though Meta has not confirmed how Wonder fits into its app tiers. Gemini agent’s pitch brings together multi-vendor model routing, persistent coworker identities and project spending limits.

Who Should Pay Attention Right Now

Your role, workplace apps and budget determine what to check during the private beta. Here is who should pay attention first.

• Workspace administrators. Coworker agents receive accounts and identities of their own. Decide what information you will share and check how permissions, action records and spending limits apply.
• Microsoft 365 and Slack teams. Google lists both as access channels. You have reason to examine Gemini agent even if Workspace is not your main workplace.
• Finance and legal teams. Both industry editions are in preview. The finance version includes more than 50 foundational skills and information from FactSet, LSEG, S&P Global and SEC filings.
• Small businesses. Wait for general availability and the Workspace consumption terms before setting a firm budget. No published rate means you cannot yet calculate a dependable monthly charge.
• Individual users. Spark is the personal agent for consumers on Google AI subscriptions. Check its access requirements before treating a Gemini agent announcement as something included in your plan.
• Teams considering a beta. Ask your Google Cloud representative about access, confirm available spending caps and decide which data to share. If you are tracking the race to fund AI agents, those practical details deserve as much attention as the announcements.

Our Take

We think model choice and company controls make the strongest case for Gemini agent. You can use Google and Anthropic models through one work agent, with identifiable coworker accounts and recorded actions. That leaves price and reliability in everyday use as the questions to watch next.

Keep Google’s customer results in perspective, too. They are company-reported claims about its AI products and Gemini Enterprise, rather than proof of what your Gemini agent will deliver. Watch for the reported late October or early November release window, then check the published terms before committing your team.

FAQs

These answers cover the main questions about access, costs and product differences.