Google Wants an AI Coworker in Every Inbox: Inside the Gemini Agent for Work

On October 8, Google Cloud CEO Thomas Kurian walked onto a stage at the Gemini at Work 2026 conference and made a claim about the future of work in a single sentence. "Work now starts in the prompt window." The product behind the line is Gemini, a single AI agent that Google describes as universal: it answers questions, handles tasks, creates content, and writes code, all from one interface.

The agent is Google's answer to a race that intensified in 2026, with OpenAI's always-on dots agents arriving in September and Meta's personal agent Muse released earlier in the fall. Google is now making the workplace its chosen battleground, with a design decision its rivals have not made: the Gemini agent gets its own identity inside the company.

An objective, not instructions

The core pitch is that you give Gemini an objective rather than step-by-step instructions, and it figures out the steps. Google's own example: ask the agent to set up a meeting with "the usual team of regional event leads," and it works out who those people are from chat spaces and past email threads, checks their calendars, and starts an email thread to coordinate a time. Ask it to research market trends, build a financial model in Sheets, and turn both into a slide deck, and it carries the project across those tools without being re-briefed at each step.

The agent runs in Google's cloud, so it keeps persistent memory regardless of how you reach it: web, Android or iOS phone, desktop, command line, or from inside Google Workspace, Microsoft 365, and Slack. For larger jobs, Gemini creates temporary helper agents, each handling one subtask and each carrying its own identity for auditing.

The coworker with an inbox

The most striking feature is the "coworker" agent. Configured with a standing role on a team, such as event coordinator, the agent receives its own Gmail address, its own calendar, and its own file storage. Teammates can add it to a chat or tag it the way they would a colleague. Its contributions show up under its own name in a document's edit history, and it sees only what team members share with it.

That design says something about where Google thinks workplace AI is heading. Chatbots that summarize a transcript after the fact are reactive. An agent with an email address and an edit history is a participant, one that can be assigned work and held to account under its own name. Google is betting that the friction in workplace AI is not model quality but organizational identity: if agents cannot be named, limited, and logged, companies will not trust them with real work.

"Work now starts in the prompt window." Google's Thomas Kurian, unveiling an agent that gets its own email address and its own edit history.

Model-agnostic by design

Unusually, Gemini is separated from the model running underneath it. By default it picks the best model for each task, choosing for accuracy and cost, and users can override the choice. At launch the options are Google's own Gemini models and Anthropic's Claude models, with Google saying more will follow. The company argues this lets enterprises switch models later without losing data or settings.

It is a strategic admission wrapped in a feature. Google spent much of 2026 chasing rivals on model releases, and early coverage described it as late to the agent party. During the July Q2 2026 earnings call, CEO Sundar Pichai told analysts that nearly 90 percent of the Fortune 100 use Gemini Enterprise, and the plan is to convert that footprint into the operating system for workplace agents, whatever model does the thinking.

The plumbing: tools, skills, and four kinds of memory

An agent is only as useful as what it can touch. Gemini ships with connections to Slack, Teams, Microsoft Office, Workspace, and Confluence for collaboration; Git and Jira for software development; Salesforce and ServiceNow for business operations; and BigQuery, Databricks, Postgres, and Snowflake for data. Organizations can add their own connections through Model Context Protocol servers, and teams can publish custom tools and saved instructions, called skills, to a company-wide library.

Google also gave the agent four memory systems. Task memory holds what it needs to finish the job in front of it, across tasks that run for days. Knowledge memory accumulates facts from documents it reads, people it talks to, and other agents it works with. How-to memory stores knowledge of how jobs get done, including skills the agent writes for itself. History memory keeps a record of everything it has already done.

The governance layer

The announcement spent as much energy on control as on capability. Google says any company deploying agents should be able to answer four questions: who is the agent, what is it allowed to do, what did it do, and what should it never touch. Every Gemini agent gets its own identity and limited access based on its role. Every action is logged under the agent's name, not a person's, and when the agent connects to outside software it signs in as itself using the same standard login behind "Sign in with Google."

Agents run in a walled-off space, and all their traffic passes through a checkpoint called the Agent Gateway, which Google calls an AI network firewall. A company writes a rule once, for example that agents may not open documents marked need-to-know, and it applies to every agent. Administrators can also set a hard spending cap per project; when it is reached, the agent pauses until someone clicks to resume, and spending is tracked per project so costs can be billed back to departments.

What it costs, and what is next

Pricing details are still thin. A Google spokesperson told VentureBeat that Gemini will be built directly into Gemini Enterprise and available at no additional charge for customers where Gemini Enterprise is available, including through existing Google Workspace business and enterprise subscriptions.

Industry-specific versions are coming first for regulated sectors. Financial services and legal versions are now in preview, with government, healthcare, and retail editions to follow. That ordering is telling: those are the industries where agents handling sensitive records face the strictest oversight, and also where the budgets are. If agents earn trust in banking and law first, the rest of the enterprise will follow.

The open question is reliability. An agent that can plan, execute, and spend across systems is only valuable if its completion rates hold up under real permissions and real budgets. Google's task inbox lets employees monitor code execution, tool use, and delegation, so oversight is part of the product surface rather than an afterthought. Buyers will need to test end-to-end completion, permission boundaries, and total task costs before treating the rollout as solved.

What is clear is the direction of travel. The workplace chatbot era, which asked workers to describe every step, is giving way to the delegation era, which asks them to describe the goal. Google's version arrives with an inbox, an identity, and a budget cap. Now it has to prove it can be managed.