Products#Multi-agent#Agent memory
Google launches the Gemini agent for business users
Google Cloud launched the enterprise Gemini agent: give it objectives, it connects to Workspace, Jira and any MCP server, and can route work to Claude.
At its annual “Gemini at Work 2026” event on October 8, Google Cloud introduced the Gemini agent, a general-purpose office agent for businesses. In Google Cloud CEO Thomas Kurian’s framing, users hand it “objectives, not just instructions” — the agent plans its own steps, delegates subtasks to subagents, connects to the systems a company already runs, and reports back through a tasks inbox. Businesses get it first; a consumer version follows later.
Key points
- Objectives, not instructions: give it a one-line goal and it decomposes, executes and reports — the dividing line from a chatbot with tools bolted on.
- A real employee identity: each agent gets its own Workspace identity — email address, calendar awareness and an audit trail attributed to the agent itself.
- Models are swappable: Gemini picks models by default, users can override with third-party options starting with Anthropic’s Claude, with open and private models promised next.
Background
Agents plugging into enterprise systems have become this quarter’s platform battleground. OpenAI shipped Dots, its always-on cloud agents, at DevDay on September 29, while Meta’s Muse approached from the personal-life side — both covered here before. Google’s answer leans enterprise: Sundar Pichai put two numbers on the board at the event — more than 1 billion monthly Gemini users, and about 90 percent of the Fortune 100 using Gemini Enterprise. Compared with a model drop like Gemini 4 Argon, the Gemini agent pitches an operating form — inside the org, with an account, under audit — and leaves raw model intelligence out of the story.
The facts
| Item | Detail |
|---|---|
| Announcement | Google Cloud “Gemini at Work 2026”; Kurian bylined blog post and YouTube demo |
| Availability | Businesses first, consumers later |
| Connectors | Workspace, Microsoft 365, Slack, Jira, Confluence, Git, BigQuery, Databricks, Postgres, Snowflake, plus any MCP server |
| Identity | Standalone Workspace account: email, calendar awareness, audit trail under the agent’s name |
| Memory | Per QbitAI: session, semantic, procedural and episodic memory |
| Model routing | Gemini auto-selects by default; Claude is the first third-party option; Smart Routing keeps context and data across models |
| Spend controls | Multi-model orchestration, smart routing, real-time spend caps |
| Clients | iOS, Android, Windows, Mac, CLI, Workspace, Microsoft 365, ServiceNow, Slack |
The customer list is worth a scan: TechCrunch reports Shopify and PayPal are testing the agent, and the published Gemini Enterprise roster includes BNP Paribas, Bradesco, Merck, Orange Spain, Santee Cooper, SOMPO, Ulta Beauty and Wesfarmers. QbitAI adds one product detail with real weight: you can create a “Coworker Agent” — onboarded like a new hire, with its own account, calendar and long-running assignments. That is the productized version of the standalone identity.
What others say
TechCrunch quotes Kurian directly — users give it objectives, not just instructions — and notes that the flexible spending options (multi-model orchestration, smart routing, real-time caps) show Google preparing to bill for a single task that crosses several models. QbitAI places the Gemini agent on the same competitive map as Meta’s Muse and OpenAI’s Dots, betting on office work, personal life and always-on cloud respectively. Google’s own blog leans hard on trust: auditable tasks and permissions that follow identity. No pricing was announced, and the consumer version has no date.
Our take
Handing agents a corporate identity plus an audit trail is the single heaviest move here: the old blocker for enterprise adoption was never capability, it was accountability — who is responsible when the agent errs. Whoever operationalizes that model first inherits the budget. For Anthropic, being a routable model inside Google cuts both ways: distribution gained, entry point ceded; commoditization of the model layer moves another step. For developers, full MCP compatibility means toolchains you have already written can attach directly — the shortest path to trying this today. Honest uncertainty: pricing is unannounced, the terms of “businesses first” are unclear, and the consumer launch is a “later” without a date.