Google announces Gemini work agent with cross-app execution and enterprise controls

Google announced its Gemini work agent at Gemini at Work 2026 on October 8, aiming to bring questions, knowledge work, content creation and coding into a single task interface. The company says users can delegate an objective and let the agent call tools, connect to business systems and carry out the work. The announcement centres on integrating workflows with enterprise controls; it does not establish that every feature is generally available to every customer. ¹

Office-setting file image from Google’s official image library. Source: Google.

Connecting applications into continuing tasks

Google Cloud CEO Thomas Kurian’s presentation draws specific boundaries around application support. Gemini can work directly inside Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar. Microsoft 365 and Slack are listed as access channels or connected systems, which does not establish equivalent inline capabilities in every external application. Google describes an agent that keeps running in the cloud, retains memory and context across devices, and can enlist sub-agents for multistep tasks. ²

The design addresses the handoffs that make office automation difficult. Retrieving information, preparing documents and running code can each be handled separately, but a business task often needs information to pass between them. A common agent could reduce repeated briefing. In return, organisations need clearer boundaries around which information enters shared context and which actions still require human approval.

Google also described coworker agents with their own identities, Workspace accounts, email, calendars and Drive, accessing only material shared with them. That creates an important accountability distinction from software operating as an individual employee: an audit needs to identify which agent acted and what permissions it inherited. ²

Official file portrait of Google Cloud CEO Thomas Kurian. Source: Google.

The work agent and its underlying model are separate choices

Gemini is both a product interface and an orchestration layer. Google says it can currently select between its Gemini model family and Anthropic’s Claude models; support for other private and open models is planned for the future. The model can change with the task without requiring the organisation’s working context to move with it. ²

Architecturally, this broadens the competitive question from the quality of an individual answer to the ability to organise an entire task. Procurement assessments need to consider model performance, the quality of tool connections and recovery after a task fails. Model choice alone does not guarantee lower bills. Routing can produce measurable savings only when complexity, retries and execution time are also managed.

Permissions, auditing and budgets in one design

Google’s governance description includes agent identities, administrator-approved granular permissions, agent-attributed audit trails, Agent Sandbox and Agent Gateway. The sandbox provides an isolated execution environment; the gateway controls network traffic under organisational policies. Google also described hard project spending limits that pause an agent when its budget is reached, leaving administrators the option to resume it. ²

These mechanisms address practical deployment problems. Generating an answer and safely changing business information require different levels of assurance; a traceable log does not make every action correct. Results still depend on whether permission settings, exception handling and human review cover high-risk steps. The announced safeguards are features to test, rather than proof of zero risk.

Preview, roadmap and availability remain distinct

Industry versions are at different stages. Financial-services and legal capabilities are in preview, while government, healthcare and retail applications are on the roadmap. The announcement provides no comprehensive timetable covering all product tiers, regions and connectors. The entire package therefore should not be presented as an already universally available service. ²

Google’s announcement points towards enterprise AI that executes continuing work rather than offering isolated functions. The useful comparison ahead is how many verifiable tasks customers can complete within defined permissions and budgets, and whether error rates, review workload and total costs improve. The customer outcomes Google presents remain company-reported experience. They cannot replace independent testing or establish how the new agent will perform in every organisation. ²

Source notes

1. 2026-10-08 · Google Cloud introduces the Gemini agent. · Google ↩

2. 2026-10-09 · Welcome to Gemini at Work 2026: Introducing the Gemini agent · Google Cloud, Thomas Kurian ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶