Gemini Agent for Work: Google Cloud's Smart, Surprising Bet

The Gemini agent is Google Cloud’s new single AI agent for work, announced on Thursday 8 October 2026 at the company’s Gemini at Work 2026 event. Google describes it as “a universal agent for work” that answers questions, handles knowledge work, creates images and media, and writes and runs code, all from one prompt box. Reuters framed the launch as Google’s answer to a crowded month, in which OpenAI shipped its Dots agents and Meta released Muse.

The pitch from Google Cloud chief executive Thomas Kurian is simple: “You give it objectives, not instructions. You delegate an outcome and come back to finished work.” The Gemini agent plans a job, picks tools and skills, connects to company systems and returns something finished inside Gmail, Docs, the inbox or a developer environment. It runs on Google’s own Gemini models and on Anthropic’s Claude models, and it is in private preview today.

This article explains what Google announced, how the Gemini agent works, what it will cost, how it compares with Microsoft, OpenAI, Meta, Anthropic and xAI, and what UK businesses should check before a pilot. Earlier today we covered a related Google change, the Antigravity Chief of Staff rename, which points in the same direction.

Table of contents

What Google Cloud Announced With the Gemini Agent
How the Gemini Agent Works Under the Hood
The Gemini Agent Inside Google Workspace
Model Choice: Why the Gemini Agent Runs Claude Too
Gemini Agent Security and Governance Controls
What the Gemini Agent Will Cost
Data, Industry and Infrastructure Announcements
The AI Agent Race the Gemini Agent Joins
Customer Results Google Cited for Gemini Enterprise
What the Gemini Agent Means for UK Businesses
Open Questions About the Gemini Agent
Frequently Asked Questions About the Gemini Agent
References

What Google Cloud Announced With the Gemini Agent

Kurian set out the launch in a keynote blog post on the Google Cloud blog. The headline product is the Gemini agent, but the post bundles seven announcements: the agent itself, Gemini inside Google Workspace, new data and analytics skills, industry versions for financial services and legal teams, agent security controls, cost controls, and a long list of customer results.

Google opened with scale figures. In the last year, nearly 500 Google Cloud customers each processed more than one trillion tokens. Nearly 80% of all Google Cloud customers use its AI products, and nearly 90% of the Fortune 100 use Gemini Enterprise. “Organizations have moved past experimentation and are running their business on it,” Kurian wrote.

One agent, one prompt box, one API

Google’s central claim is that one Gemini agent replaces a shelf of separate assistants. It answers questions in chat, works on its own towards objectives you assign, and generates code from the same interface. You can assign it work, schedule tasks or have it respond to events. Everything sits behind “a single agent and a single API”.

Personal assistant, team member or role

The Gemini agent can work for one person, for a group “like a project manager within a team”, or for a specific role “like an analyst in your finance department”. That last mode matters most for businesses, because it turns the agent from a personal tool into something closer to a member of staff.

Where the Gemini agent is available today

According to 9to5Google, the Gemini agent is in private preview, with wide availability promised “soon” for Workspace customers on select Business and Enterprise plans. Google has not published a general availability date. Alphabet shares were up marginally in early trading after the news, Reuters reported.

How the Gemini Agent Works Under the Hood

Kurian’s post lists six design principles: a unified agent, access from anywhere, persistent execution, multi-agent orchestration, deep context and model choice. Together they describe an agent that lives in Google’s cloud rather than on a laptop.

Access from any device or channel

The Gemini agent can be reached on the web, on iOS and Android, on Windows and Mac desktops, and through channels such as the command line, Google Workspace, Microsoft 365 and Slack. It can also run as a “headless agent” inside third-party applications, with no interface of its own. In Gmail, Docs, Sheets, Slides and Chat, users can type @Gemini to call it.

Persistent execution in the cloud

Because the agent runs in Google Cloud, it keeps “a single set of memories, context, and one personalization graph” whichever device you use. Work that takes hours or days keeps running after you close your laptop. Google’s Gemini for business page adds that lengthy jobs “run on their own secure machine in Google Cloud”.

Sub-agents and coworker agents

The Gemini agent “doesn’t work alone”. It can create a roster of temporary, job-specific sub-agents, each with its own identity, and coordinate them through parallel and sequential steps that can run “for hours or days”. It can also act as a longer-lived coworker agent with a defined role. Coworker agents get their own @agents.company.com email addresses and persistent storage, and they see only the context that you or your team give them.

Four kinds of memory

Google says the Gemini agent “onboards itself the way a new hire would”. It keeps four kinds of memory, summarised below from Kurian’s post.

Memory type	What it holds	Why it matters at work
Session	The task in front of it, even when it runs for days	Long jobs survive a closed laptop
Semantic	A structured knowledge base built from documents, people and other agents	No need to re-brief it on the business
Procedural	How a job gets done, including skills it writes for itself	Repeat work should get more consistent
Episodic	Everything it has done before	A record that can be reviewed and audited

Skills, tools and the company registries

Three things feed the agent. Tools connect it to Confluence, Microsoft Office, Teams, Slack, Workspace, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, Snowflake and desktop files, plus any Model Context Protocol (MCP) server. Skills are reusable instruction sets stored as modular prompts. Context is the memory above. Teams can publish tools and skills to shared company registries, building on the skills that replaced Gems last month and the custom MCP server support added to Gemini Business in September.

The Gemini Agent Inside Google Workspace

Google says the Gemini agent works “inline: in the email thread, in the document, and in the chat space”, across Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, carrying the same memory, skills and controls everywhere. It works in three ways.

Personal assistance

Kurian’s example is a meeting request. You ask Gemini to set up a meeting with “the usual team of regional event leads” next week without naming anyone. Gemini works out who they are from your chat space and the thread from your last event, checks calendars and starts an email thread to agree a time, even with external guests.

Proactive delegation

Workspace Intelligence spots work the Gemini agent could take on. If your manager emails asking for a project update as a slide deck, Workspace offers a single-click option to hand it to Gemini. In the inbox, it surfaces “the message that matters most rather than the one that arrived last” and explains why.

A coworker agent with its own account

You describe a role and Gemini creates a coworker agent for the whole team. It receives its own Workspace account, with an email address, calendar, Drive and a place in the company directory. Colleagues add it to a Chat space or @mention it. It acts “under its own identity rather than yours”, appears under its own name in version history, and “sees only what you share with it”.

VentureBeat notes that Google has not said whether every Gemini agent needs a separate account, how licensing for these accounts will work, or whether admins can switch off parts of an agent’s Workspace presence. Our recent piece on the AI coworker wave covers why named agents on the org chart raise oversight questions.

Model Choice: Why the Gemini Agent Runs Claude Too

The most surprising detail is that Google’s agent is not tied to Google’s models. “Gemini is the agent, and the model underneath it is a separate choice,” Kurian wrote. Each job runs on the model that fits best, chosen from the Gemini family and Anthropic’s Claude models today, with “other leading private and open models in the future”.

Argon, Flash, Omni, Gemma and Claude

Google’s own line-up has four tiers: Argon for frontier reasoning, Flash for speed and volume, Omni for generative media and Gemma for lightweight, open-weights edge workloads. According to SiliconANGLE, simple tasks might run on Flash while long-horizon work runs on Argon, with Claude available alongside.

Why Google is opening the model layer

Kurian’s reasoning is commercial. “The best model for the task is not always the largest one,” he wrote, and “the leading model changes every few months”. Keeping the choice open means “your context, your skills, and your data stay put” when a better model arrives. Google cited PayPal routing 10 million multi-model requests a week and Shopify blending frontier models as proof that this works at scale.

What analysts make of it

Holger Mueller of Constellation Research called the open approach the standout feature, because it “allows enterprises to use already trusted and purchased LLMs as part of the automation scope of Gemini agent”. VentureBeat adds that Microsoft also supports several models, so this is a new front in the race rather than something only Google offers.

Gemini Agent Security and Governance Controls

Kurian says two things decide whether an agent programme succeeds or stalls: “whether you can govern it, and whether you can afford it.” On governance, Google frames the problem as four questions, each answered by a named control.

Google’s question	Control	What Google says it does
Who is the agent?	Identity	Each agent gets a cryptographically attested identity with least-privilege permissions
What may it do?	Authorization	Role-based permissions approved by security admins, propagated through OAuth
What did it do?	Auditing	Every action is logged and attributed to the agent, not a person
What should it never touch?	Agent Gateway	An AI network firewall that enforces one policy across every agent

Identity, authorization and audit

The agent’s identity is stamped into the logs that capture its work and into any virtual machine it starts to run code. Because actions are attributed to the agent rather than to a person, security teams can watch those logs in real time. This is a cleaner model than agents that borrow a human’s full set of permissions, and it maps onto how cybersecurity teams already manage service accounts.

Agent Sandbox and Agent Gateway

Every Gemini agent runs tasks inside an Agent Sandbox with its own network boundary. All traffic in, out and between agents passes through Agent Gateway. You write a policy once, such as “agents may not open documents classified Need to Know”, and it applies to every agent in the company. For UK firms, these controls should sit inside an existing IT governance framework rather than beside it.

What the Gemini Agent Will Cost

Google has not announced a separate price for the Gemini agent. VentureBeat reports that Google has not said whether every capability will be included in existing Gemini Enterprise subscriptions. What Google did announce are cost controls, framed against a striking statistic: per-token prices have dropped 98% since 2024, yet enterprise AI volume “has exploded”.

Multi-model orchestration and Smart Routing

The first lever is model choice. The Gemini agent can split a large project across models, sending simple loops to cheap models and hard reasoning to frontier ones. Smart Routing does this automatically, triaging each workload “so each one runs on the model that delivers maximum performance at the lowest possible cost”.

Real-time spend caps

You can set a hard limit on a project’s AI spend in the Cloud Billing Console. The Gemini agent tracks token use and sandbox costs, and if the cap is hit, that project’s agent pauses until someone resumes it with one click. Because tracking is per project, finance teams can charge AI costs back to departments.

Savings plans and deferred execution

Spend caps are not new. Google’s late-August billing update introduced them alongside alerts at 50%, 80% and 100% of a budget, a pay-as-you-go edition of the Gemini Enterprise app, and Flexible Savings Plans. The chart below shows the discounts Google has published for agent workloads.

Discounts Google has announced for Gemini Enterprise workloads

Deferred execution, off-peak (coming soon), up to 50%

Flexible Savings Plan, 3-year commitment 20%

Flexible Savings Plan, 1-year commitment 10%

Deferred execution, which Google says will let eligible jobs run in off-peak windows for “up to half the inference cost”, suits agent work that can wait overnight. The savings plans have no minimum or maximum spend and draw down against an existing Google Cloud agreement.

Data, Industry and Infrastructure Announcements

Alongside the Gemini agent, Google gave it skills for specific domains, starting with data. Engineers describe an outcome in plain language and Gemini writes PySpark code, provides notebooks, trains models and fixes pipeline issues. Business users can ask for an operational report; Gemini builds and saves the query in BigQuery, and saved reports then run “without incurring token costs”.

Knowledge Catalog, Smart Storage and the borderless lakehouse

Three services ground those answers. The Knowledge Catalog maps business definitions such as “net margin” once, so every agent uses them. Smart Storage enriches unstructured files in place, since Google says 90% of enterprise data is unstructured. The borderless lakehouse lets Gemini query Amazon S3 and Azure Data Lake “with no variable egress fees” and read Salesforce Data 360, SAP, ServiceNow and Workday without copying data. Bloomberg Media lifted its SQL query accuracy by 63% by grounding its data agents in the catalog.

Versions for financial services and legal teams

Industry versions of the Gemini agent are in preview for financial services and legal, with government, healthcare and retail “coming soon”. The financial services version ships with more than 50 foundational skills and draws on FactSet, LSEG and SEC filings, showing confidence scores and source citations. CME Group and Deutsche Bank already use it. The legal version inherits matter-level permissions and ethical walls from NetDocuments and iManage, and Cooley is building a redaction agent on it.

TPU 8i and the price-performance claim

Google says its latest TPU 8i system delivers 80% better price-performance than the previous generation, which it presents as the reason the economics of long-running agents work. One edge example stood out: NASA’s Jet Propulsion Laboratory is running Gemma on a satellite in orbit, which Google calls a first for a vision-language model.

The AI Agent Race the Gemini Agent Joins

Reuters set the launch against “tech companies [racing] to sell AI that can work autonomously”. Since August every major lab has launched a long-running agent, and VentureBeat lined them up against Google’s.

Microsoft’s new Copilot

On 25 September Microsoft rebuilt Copilot around Home, Cowork, Code and an always-on Autopilot agent with its own identity, memory and workspace. We covered the new Copilot and its usage-based billing at launch. It is the closest match to Google’s vision, and Autopilot is in private preview too.

OpenAI’s Dots and Meta’s Muse

OpenAI introduced its always-on Dots agents at DevDay on 29 September, and Meta released Muse, a personal agent that can shop, book travel and make payments and is free for most uses, in early September. Our Dots vs Muse comparison explains how differently the two are priced.

Anthropic’s Claude in Workspace and xAI’s Grok Bot

Two days before Google’s event, Anthropic announced native Claude editing for Google Docs, Sheets and Slides. xAI’s Grok Bot, which VentureBeat now calls a SpaceXAI product, launched in August as persistent AI workers with their own cloud computers, adding shared Team Bots in September. The table summarises the field.

Product	Announced	Main pitch	Status
Google Gemini agent	8 Oct 2026	One universal agent, coworker agents with Workspace accounts, Gemini and Claude models	Private preview
Anthropic Claude for Workspace	6 Oct 2026	Edit Docs, Sheets and Slides from Claude or a sidebar	Announced
OpenAI Dots	29 Sep 2026	Always-on agents across connected apps	Paid plans only
Microsoft new Copilot	25 Sep 2026	Home, Cowork, Code and the Autopilot agent	Frontier programme and private preview
xAI Grok Bot and Team Bots	Aug and Sep 2026	Persistent workers with their own computers	Team Bots in public beta
Meta Muse	Sep 2026	Personal agent that shops, books and pays, free for most uses	Available, consumer

Where Google has an edge, and where it does not

Google’s advantages are distribution and data. Workspace already holds the email, files and calendars a Gemini agent needs, and the Accenture-led push we covered in our Accenture deal analysis adds 1,000 forward deployed engineers. Its weakness is timing: many UK firms already standardise on Microsoft 365, and Google’s agent is still a preview with no published price.

Customer Results Google Cited for Gemini Enterprise

Kurian’s post names more than 50 customer organisations. Their figures describe Google Cloud’s existing products, mostly Gemini Enterprise, not the new Gemini agent, which VentureBeat stresses has not yet been adopted at that scale. All of them are vendor-reported and none has been independently verified.

Banks and financial firms

BNP Paribas is rolling Gemini Enterprise into an internal assistant used by more than 65,000 employees. Bradesco cut document review time from one hour to five minutes and reduced risk inconsistencies by 60%. Commerzbank cut document quality checks from 20 hours to one. DBS runs chains of 70 to 80 specialised agents to draft corporate credit memos.

UK organisations on the list

Several UK names appear. Lloyds Banking Group’s Envoy pipeline has onboarded nearly a thousand engineers building AI use cases. Starling’s Scam Intelligence tool, which checks marketplace ads, has quadrupled the rate at which customers cancel likely fraudulent payments. Arden University is giving 30,000 students and staff access, and Ryanair is deploying to 35,000 employees.

How large the time savings are

Where Google gave a before-and-after figure for one task, the saving can be calculated. Snap cut diagnostic troubleshooting from 30 minutes to 30 seconds, a 98.3% reduction. Globo cut plan validation from 20 minutes to 6 seconds (99.5%). Commerzbank went from 20 hours to one (95.0%), Bradesco from 60 minutes to five (91.7%) and Grupo Bafar from 140 minutes to 17 per branch (87.9%).

Time cut on one task, calculated from Google’s before-and-after figures

Globo, plan validation 99.5%

Snap, diagnostic troubleshooting 98.3%

Commerzbank, document quality checks 95.0%

Bradesco, document review 91.7%

Grupo Bafar, contract preparation 87.9%

The pattern is clear: the biggest gains come from narrow, repeated document and diagnostic tasks, not from open-ended work. That is a useful guide for where to test a Gemini agent first.

What the Gemini Agent Means for UK Businesses

For most UK organisations the Gemini agent is not something to buy this week. It is a preview, Google has not priced it, and many firms run Microsoft 365 rather than Workspace. But it changes the questions an AI strategy should be asking.

Workspace firms and Microsoft 365 firms

Workspace customers are the obvious first audience, since the coworker accounts, inline help and proactive delegation all live in Google’s apps. Microsoft 365 firms are not shut out: Google says the Gemini agent works through Microsoft 365 as a channel and connects to Teams as a tool. The real choice is which company’s agent sits at the centre of your work, and both Google and Microsoft now want that role.

Questions to ask before a Gemini agent pilot

Before any pilot, get written answers to five questions. Which data will the agent see, and where is it processed? How are coworker accounts licensed and removed? Can admins switch off parts of the agent’s Workspace presence? Which models run which jobs, including Claude? What happens when a spend cap pauses a job halfway through?

Agents that read customer records also raise UK GDPR duties. The ICO’s guidance on AI and data protection is the starting point for a data protection impact assessment.

A 30-day evaluation plan

Pick one narrow, repeated task with a clear before-and-after time, like the document reviews in the chart above. Set a hard project spend cap from day one. Give the agent the minimum data it needs, and review its audit log weekly. At the end, compare time saved against cost and error rate. Our intelligent automation team uses the same pattern for agent pilots.

Open Questions About the Gemini Agent

Google’s announcement is detailed on architecture and thin on commercial terms. The table separates what Google has said from what it has not.

Topic	What Google has said	What it has not said
Availability	Private preview; wide availability soon on select Workspace plans	A general availability date or full rollout plan
Price	Spend caps, Smart Routing, savings plans	A price for the agent or which plans include it
Coworker accounts	Own email, calendar, Drive and directory entry	How the accounts are licensed or partly disabled
Reliability	Customer results for Gemini Enterprise	Independent benchmarks for long, multi-app tasks

Reliability over long tasks

VentureBeat points out that Google has supplied no independent benchmarks showing how reliably the Gemini agent completes long, cross-application jobs. An agent that runs for days can make a small early error that compounds, so audit logs and checkpoints matter more than headline speed.

Identity sprawl

Giving every coworker agent its own account solves one problem and creates another. A company with dozens of agents now has dozens of identities to provision, review and retire. That work belongs in the same joiner-mover-leaver process used for staff.

Frequently Asked Questions About the Gemini Agent

What is the Gemini agent?

The Gemini agent is Google Cloud’s single AI agent for work, announced on 8 October 2026. It answers questions, does knowledge work, creates media and writes code from one prompt box, and it can run sub-agents and team coworker agents.

Is the Gemini agent available now?

Only in private preview. Google says wide availability is coming soon for Workspace customers on select Business and Enterprise plans, with no firm date.

Does the Gemini agent use Claude?

Yes. It runs each job on the model it judges best, choosing between Google’s Gemini models and Anthropic’s Claude models today, with other models planned.

How much does the Gemini agent cost?

Google has not published a price. It has announced spend caps, Smart Routing and savings plans of 10% to 20%, but not which plans will include the agent.

Is it the same as Antigravity’s Chief of Staff?

Not as far as Google has said. Antigravity’s Chief of Staff is a developer tool feature, while the Gemini agent is the company-wide product. Both coordinate sub-agents, and Google may link them later.