Google Cloud launches a persistent Gemini agent for enterprise: coworker email, multi-day memory, Smart Routing

Google Cloud has a clear answer to the question every enterprise AI buyer is asking in late 2026: who will become the default digital coworker? On 8 October 2026, at the Gemini at Work event, the company unveiled a universal Gemini agent built for long running enterprise work. The agent can research, draft documents, write code, coordinate other agents and keep going for hours or days, while holding on to the same memory, business context and identity across apps and devices.

The announcement is not a small feature drop. Google wants Gemini Enterprise to stop being a pile of chat helpers and specialised tools, and to become one platform that can take an objective, figure out the steps, touch the right systems and hand back finished work. Employees would talk to one agent instead of juggling many. That agent would also show up where people already work: Google Workspace, Microsoft 365 and Slack, plus the web, iOS, Android, Windows, Mac and the command line.

According to reporting from VentureBeat and The Verge, the new agent is in private preview for enterprise customers. Google has not published a general availability date, a full rollout schedule or a separate price for the universal agent. Those gaps matter for buyers, and we will come back to them. First, the architecture and the competitive frame.

Why this launch matters now

The timing is not accidental. In the two weeks before Gemini at Work, several rivals moved in the same direction:

• On 25 September, Microsoft refreshed Copilot with Home, Cowork, Code and a background Autopilot agent that can keep working with its own identity and memory.
• On 29 September, OpenAI introduced Dots agents at DevDay: always on agents that keep progressing across connected applications between user messages.
• On 6 October, Anthropic announced deeper Claude integration with Google Docs, Sheets and Slides, including editing from a sidebar and from Claude's own interface.
• SpaceXAI's Grok Bot, launched in August and expanded with Team Bots in late September, already sells the idea of named persistent AI workers with their own cloud computers.

Google's message is that it will not cede the "central AI interface for work" slot. The company already says nearly 90 percent of Fortune 100 firms use Gemini Enterprise in some form, with large deployments such as BNP Paribas (more than 65,000 employees) and Japanese insurer SOMPO (more than 10,000 custom agents). Those numbers describe the existing platform, not necessarily the new universal agent, but they explain why Google feels it can push a bigger coordination layer on top.

What a universal Gemini agent actually does

At a product level, Google is selling one agent that can:

• Research information across connected systems
• Create and edit documents and presentations
• Write and run code in sandboxed environments
• Spin up temporary groups of specialised sub agents for complex jobs
• Coordinate parallel assignments and respond to schedules or events
• Keep working in the cloud while the human switches from phone to laptop

The Verge notes that users can chat with the agent and assign tasks from a single Gemini Enterprise interface, while still using Gemini inside Workspace apps such as Gmail, Drive, Docs, Sheets and Calendar. Because execution runs in the cloud, context travels with the agent, not with one device session.

Google Cloud CEO Thomas Kurian framed access through a common interface and API. The same agent identity can appear on the web, mobile, desktop OS surfaces, CLI tools, Workspace, Microsoft 365 and Slack. That is the "universal" claim in practice: one worker, many doors.

Coworker mode: agents with their own email

The most memorable part of the launch is the coworker configuration. Google says enterprises can give an agent its own identity inside the company's productivity stack. A manager could create something like an "Event Planner Agent" with:

• Its own Workspace account
• An email address in the company style (The Verge cites addresses under `@agents.company.com`)
• Its own calendar and Drive storage
• A presence in the company directory

Employees could then assign work the way they assign work to a human: mention the agent in Google Chat, request edits on a shared doc, drop tasks on its calendar. Google is careful to say that not every Gemini agent instance must get a full Workspace account. The dedicated account is one configuration, aimed at shared team responsibilities rather than one off personal assistants.

Security claims sit next to that story. Google says each agent gets a cryptographically attested identity and permissions managed through enterprise security systems. Actions land in audit logs, including when the agent creates isolated environments to run code. An Agent Gateway enforces organisation wide policies on traffic between agents and external systems. Sandboxing limits what code execution can touch.

Crucially, coworker agents are described as able to access only information that team members explicitly share with them. They do not inherit a human employee's full permission set by default. That distinction will matter to legal, finance and security teams who otherwise fear an always on agent quietly reading every mailbox and folder.

Google has not spelled out, at least in the materials covered by VentureBeat, exactly how administrators enable or disable pieces of an agent's Workspace presence, or how licensing and provisioning will work for permanent digital teammates versus short lived helpers. Those operational details will decide whether the coworker pitch is easy to adopt or stuck in IT review for months.

Memory that lasts longer than a chat window

Chatbots forget. Google is pitching agents that remember. The memory design splits into four categories:

1. Session memory for the current job 2. Semantic memory for accumulated company knowledge 3. Procedural memory for learned workflows 4. Episodic memory for previous assignments and outcomes

The idea is simple and ambitious: after enough work, the agent should understand how the company operates, who owns what, and how recurring processes usually run, so employees stop pasting the same onboarding paragraph into every prompt. Whether that memory stays accurate, auditable and free of stale or sensitive leftovers is an open systems problem. Google's announcement describes the categories. Independent evaluations of reliability on multi day, multi app tasks were not published with the launch.

Connectors, MCP and Smart Routing

An agent that only talks inside Workspace would not win enterprise deals. Google lists connectors that reach the tools companies already pay for, including Salesforce, ServiceNow, Jira, Git, BigQuery, Snowflake, Databricks and Microsoft Teams. Enterprises can also attach Model Context Protocol (MCP) servers and publish reusable instructions and tools into shared internal registries.

That combination is meant to support end to end jobs: research market trends, pull internal numbers from a warehouse, build a model, and produce a slide deck without a human directing every hop. Google has not released comparative benchmarks that show how often such long, cross application runs actually succeed without human rescue.

Cost control is the other half of the story. A new Smart Routing feature selects models per workload based on performance and price. Google supplies Gemini models, and already supports Anthropic's Claude models inside the same agent infrastructure, with more proprietary and open weight models planned. In theory, an enterprise could keep the same agent identity, company memory and reusable skills while swapping the model that performs a given step.

Google also highlights project level spending limits that can pause agent activity when a budget runs out, plus the ability to attribute spend to departments. Hard project spend caps are not brand new (Google talked about them in August, along with pay as you go and commitment discounts), but tying them to autonomous agents that can burn tokens overnight is a practical necessity. What Google has not done is announce a separate list price for the universal agent, or say clearly which capabilities sit inside existing Gemini Enterprise subscriptions versus add ons.

Specialised industry agents on the side

Alongside the general purpose agent, Google is shipping specialised skill packs. Financial Services, currently in preview, includes more than 50 foundational skills and links to providers such as FactSet and LSEG. A Legal offering focuses on document management permissions and confidentiality, with support for platforms including NetDocuments and iManage. Google also points to specialised capabilities for data engineering and business analytics.

These vertical agents matter because many buyers do not want a free floating generalist with access to everything. They want a constrained worker that knows industry workflows and stays inside regulated boundaries. Still, the universal agent is the strategic centre of the announcement: the vertical packs are accelerators on top of that coordination layer.

How Google compares with Copilot, Dots, Claude and Grok Bot

The competitive table is getting crowded, and the differences are more about architecture and trust than about raw model IQ alone.

Microsoft Copilot is the closest peer. Home combines chat with delegated Cowork, Code lets employees build apps and automations, and Autopilot is the background agent with its own identity and workspace. Home and Code are rolling through Microsoft's Frontier program; Autopilot is expanding in private preview. Microsoft's advantage is deep Microsoft 365 context. Google's counter is cross platform reach (including Microsoft 365 and Slack) plus multi model support that already includes Claude.

OpenAI Dots push continuous work across connected apps and between user interactions, with ChatGPT Space for shared project context. OpenAI's strength is the ChatGPT user base and a fast agent roadmap. Google's pitch leans harder on enterprise identity, auditability and Workspace native coworker accounts.

Anthropic is taking a different path: put Claude inside the productivity apps people already open, with native Google Workspace editing and connectors that let Claude modify Docs, Sheets and Slides from Claude itself. That is less "new digital employee" and more "sharp co editor in the sidebar." Google's universal agent is broader in scope, but Anthropic's Workspace move (announced two days before Gemini at Work) undercuts any claim that only Google can live inside Docs.

SpaceXAI Grok Bot is the clearest twin of Google's coworker idea. Bots get their own cloud computers, keep running when the user's laptop is off, learn routines and coordinate. Team Bots add shared organisational knowledge with private individual conversations. One technical contrast VentureBeat notes: Grok Bot emphasises operating software through a dedicated computer environment, including apps without good APIs, while Google emphasises integrations, reusable skills, agent identities and enterprise security controls.

Google's willingness to run Claude models inside its own agent stack is a notable flex. It tells buyers that the platform layer (memory, connectors, policy, routing) can outlive any single model vendor. Microsoft is also pursuing multi model selection, so this is becoming table stakes rather than a Google exclusive.

What is still unclear

A careful reading of the launch coverage leaves several buyer questions unanswered:

• Availability: private preview for enterprise only. No GA date, no public timeline for when every promised surface (coworker accounts, full cross platform features, Smart Routing everywhere) will land for typical customers.
• Pricing: no separate price for the universal agent. Unclear which pieces are included in current Gemini Enterprise seats.
• Administration: incomplete public detail on selective enablement of Workspace identity components, licensing for permanent digital coworkers, and how permissions interact with existing identity providers.
• Reliability: no independent benchmarks on multi day, multi application task completion rates, recovery from failure, or hallucination risk when the agent writes code or edits shared files.
• Consumer path: Google previously launched a consumer facing Gemini Spark agent (May) that works across Workspace and the web through a separate Gemini app interface. The enterprise universal agent is a different product track; The Verge notes it may preview Google's broader "one window that does the work" vision, but it is not Spark.

Those gaps do not make the announcement empty. They do mean CIOs should treat the launch as an architecture preview plus a competitive signal, not as a ready to buy SKU sheet.

What enterprises should test in private preview

If your organisation already uses Gemini Enterprise, the private preview is an opportunity to pressure test a few concrete scenarios rather than a vague "try agents" goal:

1. Multi day research to deliverable: ask the agent to gather internal and external data, keep context overnight, and return a finished brief or deck without re explaining the brief each morning. 2. Coworker identity drill: provision one shared agent with email and Drive, assign work via Chat mentions, and check whether audit logs and share boundaries behave as security expects. 3. Cross stack connectors: run a job that needs Salesforce or ServiceNow plus BigQuery or Snowflake, and measure how often the agent needs human unblocking. 4. Smart Routing cost: force a mix of trivial and hard steps, then verify that spend attribution and project caps actually stop runaway runs. 5. Multi model switch: if Claude is enabled in your tenant, compare Gemini only versus Gemini plus Claude on the same workflow, keeping memory and tools constant.

Also plan the boring controls early: who can spawn coworker agents, which data classes may never be shared with them, how to revoke an agent that has its own mailbox, and what happens to that mailbox when the project ends.

The bigger shift: from chatbots to digital workers

Zoom out, and Google's launch is one more data point in a market turn that started years ago with copilots and is now landing on persistent workers. Chat UIs are still the on ramp, but the product that vendors want to own is the long lived process owner: the thing that carries context, holds credentials (carefully), talks to systems of record and returns outcomes instead of paragraphs.

That shift raises governance stakes. A chatbot that hallucinates a sentence is annoying. An agent that hallucinates a code change, a calendar invite to a customer or a Salesforce update is an incident. Google is answering with attested identities, gateways, sandboxes, explicit sharing and spend caps. Competitors are answering with their own governance stories. The winner will not only be the firm with the best model on a leaderboard. It will be the firm enterprises trust to keep working when nobody is watching the screen.

For Google specifically, the Gemini agent announcement closes a narrative hole. The company has strong models, deep Workspace distribution and a large Gemini Enterprise footprint. What it needed to show on stage was a coherent story for long running, cross application work that does not force every employee into a new silo app. Private preview status means the story is still ahead of broad deployment. The competitive clock, set by Microsoft, OpenAI, Anthropic and SpaceXAI in the same fortnight, is already running.

Sources and further reading

Primary coverage of the Gemini at Work 2026 enterprise agent launch:

• VentureBeat: Google Cloud unveils persistent Gemini Agents for long running tasks
• The Verge: Google is launching a one stop Gemini agent for your work tasks

Related competitive context mentioned in those reports includes Microsoft's Copilot Autopilot updates (25 September 2026), OpenAI's Dots agents at DevDay (29 September 2026), Anthropic's Claude for Google Workspace integration (6 October 2026) and SpaceXAI's Grok Bot / Team Bots line. Google's own Gemini Enterprise adoption claims (Fortune 100 usage, BNP Paribas, SOMPO) come from the company via that same launch coverage and should be read as vendor reported figures for the existing platform, not as proof of universal agent uptake.

In short: Google has put a serious universal agent on the enterprise table, with coworker identities, four layer memory, broad connectors, Smart Routing across Gemini and Claude, and private preview access today. The missing pieces are the GA calendar, the price sheet and independent proof that multi day digital coworkers stay reliable and safe at scale. Those are the questions buyers should bring into every Google Cloud conversation for the rest of 2026.