Google’s new Gemini agent is designed to keep working in the cloud for hours or days, carrying the same memory and context across devices and applications. An employee could assign work from one interface, close their laptop, and return through another channel without restarting the task or briefing a separate assistant.
Announced at Gemini at Work on October 8, 2026, the agent brings together chat, knowledge work, media creation, code generation and execution, scheduled tasks, and event-driven work. Google describes it as a single universal agent, accessible through one interface and one API.
That is the architectural pitch, not a declaration of broad availability. The Verge reports that the enterprise-focused agent is currently available in private preview. Google’s announcement discusses early testers but does not clearly establish a general-availability date.
The consequential change is the unit of enterprise AI Google wants companies to adopt: a persistent agent that coordinates work across systems, rather than a collection of assistants that each hold a fragment of the task.
One Agent Across Chat, Code, and Background Work
Google says Gemini can answer questions interactively, work autonomously toward assigned objectives, and write and run code through the same interface. Users can also schedule tasks or have the agent respond to events.
Those are different operating modes. Answering a question ends with a response; completing an objective may require a plan, access to external systems, tool execution, and follow-up work. Bringing them together means a conversation could become a longer-running task without moving to a separate automation product.
Google’s launch summary describes the intended output as finished work delivered inside documents, inboxes, and developer environments. That ambition goes beyond generating text for someone to copy elsewhere.
Access is similarly broad in the announced design. Google lists the web, iOS and Android, Windows and Mac desktops, command-line access, Workspace, Microsoft 365, and Slack. The agent can also operate without a dedicated user interface, allowing integration into other applications.
Inside Workspace, Google says Gemini will work directly in Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar while carrying the same memory, skills, and controls.
The important distinction is between many access points and many separate agents. Google proposes that these surfaces reach the same cloud-based agent. Whether every surface supports the same actions, and which integrations preview customers can actually use, still requires product-specific confirmation.
Persistent Memory Becomes Part of the Enterprise Architecture
Google describes a single set of memories, shared context, and a personalization graph that follows the agent across devices and channels. Work can continue after the user disconnects.
That design addresses a practical limitation of session-based assistants: a task’s context can become tied to the conversation or application where it began. Persistent execution instead makes the cloud agent responsible for carrying the work forward.
Google divides its memory architecture into four categories:
- Session memory retains context for the current task, including work that runs for days.
- Semantic memory holds structured knowledge gathered from documents, people, and other agents.
- Procedural memory records how work gets done, including skills the agent creates.
- Episodic memory records previous activity.
The distinctions matter because remembering a preference is different from retaining business knowledge or reusing a procedure that changes records in another system.
Google also describes dedicated projects that help teams organize the context, skills, and tools available to Gemini. Skills are reusable instructions, knowledge, or workflows stored as modular prompts. Teams can publish them to shared registries, while individuals can maintain personal skills.
For technical leaders, memory introduces governance questions alongside the convenience. How is an incorrect memory corrected? What happens when someone’s access changes? Which information belongs to an individual, a project, or the organization?
The announced memory categories explain what Google intends to retain. They do not, by themselves, answer those lifecycle and access-control questions.
Sub-Agents Handle Jobs, While Coworker Agents Hold Roles
The universal agent is not necessarily a single execution process doing everything itself. Google says Gemini can dynamically create temporary, job-specific sub-agents, each with its own identity, and coordinate their work in parallel or sequentially.
A multi-step assignment could therefore involve several delegated workers rather than one uninterrupted model conversation. Google says these workflows can run for hours or days.
Google separately describes “coworker agents” with persistent roles and operational presence across sessions. These agents can have dedicated identities, email addresses, and storage. According to the announcement, they receive only the context that users or team members provide.
The distinction is useful. A temporary sub-agent exists to complete part of a job. A coworker agent represents an ongoing organizational role, potentially managing changing responsibilities over time.
Neither arrangement automatically establishes reliable execution. Delegation creates additional points where instructions can be misunderstood, results can conflict, or work can fail midway.
For an enterprise evaluation, the useful question is not simply how many agents Gemini can create. It is whether the system makes responsibilities, permissions, intermediate results, and failures inspectable. A polished final answer would not be enough evidence that a complex workflow completed correctly.
Gemini Is the Agent, Not Necessarily the Underlying Model
One of the announcement’s most important distinctions is Google’s separation of the agent from the model powering a particular task.
Google says Gemini can orchestrate work across its own Gemini models and Anthropic’s Claude models. Support for other leading private and open models is described as a future direction, rather than current availability.
This architecture treats the persistent agent as the layer holding context, skills, and connections, while the model underneath can change. In principle, an organization would not need to rebuild those resources every time it selected a different model.
Google argues that routing work to an appropriate model can improve quality and reduce costs because the largest model is not always necessary. It also names Smart Routing and real-time spend caps among its cost-control features.
Those are vendor claims about the system’s intended behavior. The announcement’s architecture does not establish that routing will consistently select the best model for every enterprise task, or that it will produce a particular saving.
Google says sportswear company On tested the new dynamic-selection capability. That is relevant early-testing evidence, but it is not an independent comparison of routing accuracy, cost, latency, or end-to-end task success.
Buyers should evaluate routing against their own workloads. A cheaper model choice provides little benefit if it causes additional retries or produces work that requires substantial correction.
Cross-System Reach Makes Permissions More Important
Google’s tools layer reaches beyond its own applications. The announcement lists collaboration products, development tools, enterprise platforms, databases, and local files.
Named systems include Microsoft Office and Teams, Slack, Confluence, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, and Snowflake. Google also says Gemini can connect to Model Context Protocol servers inside or outside a company’s network.
Sources
- Gemini agentcloud.google.com
- The Verge reportstheverge.com
- launch summaryblog.google
