Google introduces a persistent Gemini work agent with permissions and spend controls

Google Cloud has announced a Gemini agent that combines question answering, knowledge work, media creation and code execution within one interface. At Gemini at Work 2026, CEO Thomas Kurian described a cloud-running agent that can accept objectives, use company tools and continue working after the user closes a laptop.

The Google Cloud announcement is adapted from Kurian's keynote. It presents Google's product architecture and customer examples, rather than an independent evaluation of reliability. The company describes an agent spanning several models and applications; the Gemini name here refers to the agent experience as well as the underlying model family.

Persistent work and distinct agent identities

Google says the agent maintains memory and context across devices and channels. It can respond to events, take assigned work or handle scheduled tasks, with long-running jobs continuing in the cloud. The described access points include browsers, mobile and desktop devices, command-line tools and collaboration applications.

For complex work, Gemini can create temporary subagents with individual identities. Google distinguishes those job-specific helpers from persistent coworker agents, which have defined roles, storage and their own company email addresses. A coworker agent sees context that the team provides and acts under its own identity, according to the announcement.

Inside Workspace, Google describes assistance in Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar. Its examples include coordinating a meeting from existing team context and creating a presentation from research and a financial model. These illustrate the proposed workflow. They do not establish that every organization can enable every example without checking access and deployment conditions.

Tools, skills and business definitions provide context

The agent connects to company systems through tools and reusable skills. Google lists collaboration services, development platforms, databases and MCP servers, alongside registries where organizations can publish tools and skills. Skills contain instructions or workflows for particular tasks.

Google describes four forms of memory: current-session context, semantic knowledge, procedural knowledge and records of earlier activity. That architecture is intended to reduce repeated briefing. Organizations still need to decide which information an agent can retain and which sources it may access; persistent context makes those decisions more consequential.

For data work, the announcement connects Gemini to Knowledge Catalog definitions and operational reporting. Google says the agent can generate SQL, Spark or Python code and run it in BigQuery or Managed Spark. A saved query can be rerun without another model-generated query. The consistency of that execution depends on checking the original query and its business definitions.

Financial Services and Legal specializations are in preview. Government, Healthcare and Retail are described as coming later. Those availability distinctions should remain visible instead of presenting the entire industry list as a finished rollout.

Governance and spending controls accompany delegation

Google says agents receive individual identities, role-based permissions and action logs. Its Agent Sandbox and Agent Gateway are intended to enforce network and organizational policies. These are announced controls, not a guarantee that an organization has already configured them correctly or that agent mistakes are impossible.

The model choice remains separate from the agent. Google describes orchestration across Gemini models and Anthropic's Claude today, with other models planned. Smart Routing is intended to select an appropriate model for a workload. Keeping context and tools separate from model selection could help teams change models without rebuilding the entire workflow.

A project-level spending cap adds another stopping condition. Google says an agent pauses when the cap is triggered, accounting for token usage and sandbox costs, and a user can resume it through the console. That gives administrators a concrete control to test alongside permissions and audit records before delegating longer jobs.