On 8 October 2026, Google Cloud chief executive Thomas Kurian took the stage at the Gemini at Work 2026 event and unveiled the Gemini agent — a “universal AI agent for work” that redefines what office software is supposed to do. His pitch line captured the whole idea in a single sentence: “You give it objectives, not just instructions.” In this guide, we explain exactly what the Gemini agent is, how it slots into Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, what it costs, who can actually get it, and how it compares with rivals such as OpenAI Dots and Claude in Chrome.
For the past three years, AI assistants have mostly waited for instructions: you type a prompt, you get an answer, you type the next prompt. The Gemini agent represents Google’s bet that the next leap looks very different — software that accepts an objective, plans the work itself, and keeps going until the job is finished, even if that takes hours or days. Whether the promise survives contact with real offices remains to be seen, but the ambition is unmistakable: Google wants its agent to become the connective tissue of the working day.

In this guide:
- What Is the Gemini Agent?
- Why Google Launched It Now
- How It Actually Works
- Inside Your Favourite Apps
- Four Kinds of Memory
- Co-worker Agents: An AI Teammate With Its Own Inbox
- Model Flexibility and Smart Routing
- Enterprise Security and Cost Controls
- Gemini Agent vs the Competition
- Pricing and Availability
- How to Get Access
- Frequently Asked Questions
- Our Verdict
In the sections below, we walk through everything announced so far: the capabilities, the memory system, the “co-worker” concept, the security controls, the pricing picture (or the lack of one), and the crucial difference between this enterprise product and the consumer feature with an almost identical name. Let’s dive in.
What Is the Gemini Agent?
At its core, the Gemini agent is an AI system built for knowledge work inside large organisations. It answers questions, digs through documents and data, creates images and media, and — the capability that matters most to technical teams — it writes and runs code. That last point is what separates an agent from a chatbot: it does not just suggest a spreadsheet formula, it builds the model, executes it, checks the output, and fixes its own mistakes.
The conceptual shift is the real story. Traditional software waits for step-by-step instructions; traditional chatbots wait for the next prompt. Here, you hand over an objective such as “prepare our Q4 market analysis” and the system decomposes it into research, modelling, drafting and review — deciding the order of operations itself and asking for your input only at genuine decision points. It is the difference between hiring a typist and hiring an analyst.
Google is positioning the product squarely at enterprises rather than individual consumers. The announcement came from Google Cloud, not the consumer Gemini team, and everything about the packaging — identity management, policy controls, cost caps — signals a product sold to CIOs and deployed across departments. That enterprise framing shapes every capability described below.
Why Google Launched It Now
2026 has been the year the AI industry stopped demoing chatbots and started shipping co-workers. OpenAI is pushing always-on agents with Dots, Claude in Chrome is now generally available with prompt-injection defences as a headline feature, and Microsoft Copilot is entrenched across Office workflows. Google could not afford to let the “agent” story be told without it.
But Google also holds advantages its rivals would love to buy: the daily habits of well over a billion Gmail and Docs users, and deep enterprise relationships through Google Cloud. This is the gap Google is betting its new agent fills. Rivals may own the model layer or the operating system, but Google owns the documents, the inbox and the calendar where knowledge work actually happens — and an agent that lives there starts with a home-field advantage.
There is a technology argument too. Long-running, multi-step agents only became practical once models grew reliable enough to plan, use tools and check their own work without drifting off course — and once the cost of running them fell far enough for businesses to stomach. If you would like to play with Google’s AI models yourself before committing to anything enterprise-grade, our Google Playground guide is a friendly place to start.

How It Actually Works
Strip away the keynote gloss and the Gemini agent is an orchestration layer: it takes your objective, breaks it into steps, and executes them by driving the apps and tools you already use. The headline example from Google says it all — research a market, build the financial model in Sheets, then draft the presentation — all from a single prompt. In practice, that one prompt unfolds into something like this:
- Research: temporary sub-agents fan out across Drive, the web and connected data sources, gathering market figures in parallel.
- Modelling: the agent writes code, runs it, and builds the financial model directly in Sheets — checking its own formulas as it goes.
- Drafting: it turns the findings into a slide deck, pulling charts from the model it just built.
- Review: it flags assumptions and decision points for you, rather than asking about every trivial step.
Those temporary sub-agents are worth pausing on. For multi-step jobs, the system spins up sub-agents that each carry their own identity, tackling parts of the work in parallel or in sequence before reporting back. It is genuine multi-agent orchestration, not just one model wearing different hats.
And because everything runs in the cloud, a job that takes hours or days keeps going after you close your laptop. You set the objective in the morning, shut the lid, and return to finished work — or at least to a clear summary of where things stand and what needs your call.
Inside Your Favourite Apps
The Gemini agent lives where work already happens. Inside Google’s world, that means Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar — the agent can triage your inbox, pull numbers from a spreadsheet, draft the document and book the follow-up meeting without you switching tabs. The pitch is that context follows you: start something on the web, check progress on mobile, review on desktop, with the same memory everywhere.
Crucially, it does not stop at Google’s borders. The agent also reaches into Slack, Microsoft 365 (including Teams), Salesforce, ServiceNow, Jira, Snowflake and Databricks — plus the command line and Git for engineering teams. That breadth matters because real work is messy: the sales data lives in Salesforce, the discussion in Slack, the deliverable in Slides. An agent confined to a single ecosystem would be a curiosity; one that crosses them starts to look like infrastructure.
For US companies in particular, the Microsoft 365 bridge is the detail to watch. Most large American firms live in a mixed Google-Microsoft world, and an agent that can pull a Teams thread into a Google Doc — or push a Sheets model into a Teams channel — removes one of the most persistent frictions in modern office life.
Four Kinds of Memory
Most AI tools suffer from convenient amnesia: every conversation starts from zero. Google’s answer is a four-layer memory system designed so that business context carries across apps and devices:
- Session memory — what is happening right now: the current task, the open thread, the half-finished draft.
- Semantic memory — durable facts about your business: product names, pricing tiers, org charts, key accounts.
- Procedural memory — how your team works: the approval chain for expenses, the format the board expects, the checklist before a launch.
- Episodic memory — what has happened before: last quarter’s campaign, the vendor negotiation that went sideways, the decision you reversed in March.
Together, these give the Gemini agent something close to institutional knowledge — the reason a good human assistant gets more useful every month. It also raises the stakes on governance, which is presumably why the enterprise security story (covered below) received so much keynote time.

Co-worker Agents: An AI Teammate With Its Own Inbox
The boldest idea in the announcement is the “co-worker agent”: each agent gets its own Gmail address, its own calendar, its own Drive storage and its own seat in the company directory. Colleagues do not “use” it so much as delegate to it — CC it on an email thread, invite it to a meeting, assign it a ticket in Jira — exactly as they would with a human teammate.

Think about what that normalises. Today, asking an AI for help means opening a separate tool and writing a careful prompt. In the co-worker model, the Gemini agent is simply there: looped into threads, aware of the calendar, holding the shared files. The interaction pattern shifts from “go ask the AI” to “let the team handle it” — where the team quietly includes software.
There are obvious questions — who is accountable when the agent commits to a deadline in a meeting it attended alone? — and sensible companies will want clear policies before granting an agent a directory seat. But as a vision of how humans and AI actually collaborate, it is the most thought-provoking thing Google announced all year, and the Gemini agent framing makes the intent explicit: not a tool you wield, but a colleague you brief.
Model Flexibility and Smart Routing
Here is a surprise: the Gemini agent is not locked to Google’s own models. It uses multi-model orchestration, selecting between Google models and Anthropic’s Claude models for each task, depending on which is likely to do the better job. A delicate piece of client communication might go one way; a gnarly reasoning problem might go the other.
Two mechanisms keep this practical rather than chaotic. Smart Routing makes the per-task model choice automatically, and real-time spend caps let finance teams set hard limits on what the agent is allowed to burn. For a technology that could otherwise rack up alarming API bills at 3 a.m., that cost control is not a footnote — it is a precondition for enterprise adoption.
A word on what sits underneath: industry press has reported the underlying model family as Gemini 3 Pro. Treat that as industry reporting, not a Google-confirmed fact — Google itself has been careful to talk about the agent layer rather than the base model. (For context on where Google’s models are heading, see our Gemini 4 Argon explainer — and if you are curious about the Claude side of the partnership, our Claude Haiku 5.5 guide covers Anthropic’s latest efficient model.)
The cross-vendor angle is strategically interesting. By routing work to Claude models where they fit best, Google concedes that no single lab wins every task — and turns that concession into a selling point. Enterprises get best-of-breed results without managing two vendor relationships, which is exactly the kind of pragmatism that wins CIO budgets, and a shrewd way to keep the Gemini agent credible even where Google’s own models are not the strongest.
Enterprise Security and Cost Controls
None of this works in a large company without the unglamorous machinery of trust, and Google devoted serious keynote time to it. The headline controls include identity and policy management, authorisation and permission controls that decide what the agent may touch, secure sandboxing so generated code runs somewhere safe, and network gateways governing what it may reach. Cost controls sit alongside the security stack, not apart from it.
Two extensibility points round out the picture. The agent connects to any MCP server — the open standard for plugging tools and data sources into AI systems — and it supports custom skills added by the user: reusable capabilities your team defines once and invokes forever. It also learns from each interaction, which is powerful and slightly unnerving in equal measure. The governance story has to cover not just what the agent can do today, but what it picks up tomorrow.
For regulated industries — and Google is already previewing industry-specific agents for financial services and legal, with government, healthcare and retail versions to follow — this is the section that decides deals. The Gemini agent will be bought or rejected on the strength of these controls far more than on the cleverness of its demos.
Gemini Agent vs the Competition
Google is not entering an empty field. The race to build the definitive work agent is the defining enterprise AI contest of 2026, and each rival is attacking from a different direction. Here is how the field looks from where we stand.
Gemini agent vs OpenAI Dots
OpenAI’s Dots is built around always-on agents that keep working in the background — the closest direct rival to Google’s vision. The boldest differentiator for the Gemini agent is the co-worker model: its own inbox, calendar, Drive storage and directory seat, plus genuine multi-model routing that can send a task to a Claude model when that fits best. Dots may match the persistence; it does not currently match the “hire it like a colleague” framing.
Claude in Chrome, now generally available, takes yet another angle: the browser as the workplace, with prompt-injection defences as a headline feature. That focus on safer web automation is smart, but it is narrower than Google’s cross-app, cross-vendor play. Microsoft Copilot remains the incumbent — deeply embedded where Office 365 rules — while OpenAI’s AgentKit gives developers the toolkit to build their own agents rather than buying one off the shelf. As PYMNTS reported at launch, Google is explicitly targeting the enterprise market with this product, and the Workspace footprint is its biggest weapon.
| Capability | Gemini agent | OpenAI Dots | Claude in Chrome | Microsoft Copilot |
|---|---|---|---|---|
| Announced / status | 8 Oct 2026, private preview | Not disclosed | Generally available | Not disclosed |
| Works across Google Workspace apps | Yes | Not disclosed | Not disclosed | Not disclosed |
| Reaches third-party tools (Slack, Salesforce, Jira) | Yes | Not disclosed | Not disclosed | Not disclosed |
| Long jobs keep running in the cloud | Yes | Not disclosed | Not disclosed | Not disclosed |
| Own inbox, calendar and directory seat | Yes | Not disclosed | Not disclosed | Not disclosed |
| Routes work across model vendors | Yes (Google + Claude models) | Not disclosed | Not disclosed | Not disclosed |
| Public pricing | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
Pricing and Availability
The Gemini agent is currently in private preview for select customers, available through the Gemini Enterprise app and Google Workspace. Google has not publicly disclosed pricing — there are no published tiers, per-seat figures or usage rates — and we will not invent any. Expect enterprise-style negotiated pricing when it becomes broadly available, with the real-time spend caps described above doing the day-to-day cost policing.
Not to Be Confused With the Consumer Gemini Agent
This distinction matters, so let’s be blunt: the enterprise product above is not the same as the consumer “Gemini Agent” feature inside the Gemini app, which launched in November 2025 for Google AI Ultra subscribers. That consumer feature handles multi-step browsing tasks on your behalf and asks for confirmation before sensitive actions such as sending messages or making purchases. Same family name, completely different product — one is a personal browsing helper, the other is an enterprise co-worker with its own inbox. For a deeper side-by-side, see this guide on the enterprise agent vs the consumer feature.
On the roadmap, industry-specific versions are already moving: agents tailored for financial services and legal are in preview now, with government, healthcare and retail editions coming soon. If your sector is on that list, the private preview conversation is worth having sooner rather than later.
How to Get Access
There is no public signup page, and Google has not announced a general-availability date. For now, access runs through the usual enterprise channels:
- Talk to your Google Cloud account team or request entry via the Gemini Enterprise programme — private preview seats go to select customers.
- Arrive with two or three concrete use cases. “Research a market, model it in Sheets, draft the deck” is exactly the kind of objective-led workflow Google wants to showcase, and a crisp pilot brief will move you up the queue.
- Sort out data governance early. Decide which drives, inboxes and third-party tools the agent may touch, who approves its permissions, and how its spend caps are set — before the pilot, not during it.
Realistically, the first wave will be large Workspace and Google Cloud customers with a clear multi-app workflow to automate. Smaller teams should watch the industry-specific previews — financial services and legal first — as the most likely path to broader access.
Frequently Asked Questions
Is this the same as the Gemini feature in the Gemini app?
No. The consumer feature in the Gemini app (launched November 2025 for Google AI Ultra subscribers) does multi-step browsing tasks and confirms with you before sensitive actions. The enterprise product described in this guide is a different beast entirely: it works across your company’s apps, keeps running for hours or days, and shows up in the directory like a colleague.
How much does it cost?
Google has not disclosed pricing. The product is in private preview for select customers via the Gemini Enterprise app and Workspace, and there are no public tiers or per-seat figures. Real-time spend caps are built in, so at least runaway bills are a solved problem by design.
Which apps does it work with?
Inside Google: Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar. Beyond Google: Slack, Microsoft 365 (including Teams), Salesforce, ServiceNow, Jira, Snowflake and Databricks, plus the command line and Git. It runs on web, mobile and desktop with the same memory everywhere.
Is my company’s data safe?
Google is pitching enterprise-grade controls: identity and policy management, authorisation and permission controls, secure sandboxing and network gateways. As with any cloud AI handling company data, your security team should review the data-processing terms before any rollout — the controls sound right, but trust is earned in the audit, not the keynote.
When will it be generally available?
Google has not announced a date. It remains in private preview for select customers, with industry-specific agents for financial services and legal also in preview and government, healthcare and retail editions on the way.
Our Verdict
Strip out the marketing and Google has made three genuinely interesting bets: that workers will hand objectives rather than instructions to software, that an agent deserves its own inbox and directory seat, and that no single model — not even Google’s — wins every task. Each bet is defensible, and together they add up to the most coherent enterprise agent vision anyone has shipped this year.
The honest caveats: it is private preview, pricing is undisclosed, and the demos have not yet survived contact with a messy real-world inbox. If the new agent delivers on even half of the keynote promises, though, “CC the agent” could become as ordinary as “CC the team” within a couple of years. For US companies already living in Workspace, this is the AI announcement of the autumn — and it is worth getting your use cases ready now.
