Google Cloud CEO Thomas Kurian used the Gemini at Work 2026 keynote on 8 October to introduce the Gemini agent: one agent that answers questions, does knowledge work, creates media and writes and runs code, through one interface and one API. It arrived a month after Meta’s consumer Muse and nine days after OpenAI’s dots, and it is aimed squarely at enterprises.

What is actually different

Sub-agents and coworker agents. For a multi-step job, Gemini creates temporary, job-specific sub-agents, each with its own identity, and coordinates them in parallel or in sequence over hours or days. Coworker agents are the persistent version: an agent with a defined role, its own storage, access only to the context a team gives it, and its own identity, including an @agents.company.com email address. In Google Workspace a coworker agent gets a full account with email, calendar, Drive and a directory entry, can be @mentioned in Chat, and suggests edits in Docs under its own name in version history.

Model choice, including Claude. Gemini routes each job to the model it judges best. Google says that today includes Gemini models and Anthropic’s Claude models, with other private and open models planned. A “Smart Routing” option picks the cheapest model that meets the bar.

Memory, in four kinds. Google names session, semantic, procedural (including skills the agent writes for itself) and episodic memory, with one memory and personalisation graph across devices.

Reach. Web, iOS, Android, Windows, Mac, the command line, Workspace, Microsoft 365 and Slack, plus a headless mode for embedding in other apps. Connectors cover Confluence, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Snowflake, Postgres and more, and any MCP server.

Governance. Each agent has a cryptographically attested identity with least-privilege permissions approved by security administrators; every action is logged against the agent, not a person; agents run in an Agent Sandbox with their own network boundary; and an Agent Gateway acts as a policy firewall for all agent traffic. Spend caps can pause a project’s agent when its budget is hit.

What Google did not say. The keynote post gives no pricing, no general-availability date and no regional availability for the agent itself. Launch coverage describes it as a private preview for enterprise customers. Industry versions for financial services and legal are in preview; government, healthcare and retail are “coming soon”.

Why it matters for builders

The identity model is the real news. An agent with its own mailbox, calendar, directory entry and audit trail is treated as a member of staff rather than as a feature of someone’s account, which is what security teams have been asking for: you can scope it, review its access and revoke it like a service account. It is the opposite design choice to Muse, which acts as you inside your accounts. Microsoft’s Copilot Autopilot went the same way as Google in September. Expect “does the agent have its own identity?” to become a standard procurement question.

The Claude routing is also worth noting. Google’s flagship agent product sends some work to a competitor’s models, which says a lot about how much model choice now matters to enterprise buyers, and it means data-processing reviews have to cover more than one model provider behind a single product.

For EU teams: nothing announced is generally available yet, so there is no residency or regional detail to plan against. Ask for it before any pilot.

Sources

Further reading