Google Launches Gemini Agent: Own Identity, Claude Models, Spend Caps

On 8 October 2026, at its Gemini at Work event, Google Cloud launched what it calls a single, universal agent for work. The Gemini agent takes goals rather than prompts, delegates to sub-agents, connects to any MCP server, and gets its own Workspace account with an email address and an audit trail in its own name. Google Cloud CEO Thomas Kurian said you give it "objectives, not just instructions." It launched in preview for a select set of customers, per CIO Dive, with wider availability "coming soon" but undated. I spend a lot of time on agent identity, and this is one of the first big vendor launches that make the agent a first-class account rather than a feature hanging off a human's login.

What happened

According to TechCrunch, SiliconANGLE and 9to5Google, the launch covered:

• One agent, many front doors. It is reachable from iOS, Android, Windows, Mac, the command line, Google Workspace, Microsoft 365, ServiceNow and Slack. In Gmail, Docs, Sheets, Slides and Chat you mention @Gemini. An API lets other software embed it as a "headless agent."
• Sub-agents and long jobs. It plans work, spawns sub-agents, and coordinates parallel and sequential workflows that can run for hours or days. A "tasks inbox" shows reasoning, delegation, loaded skills, code and progress.
• Its own identity. The agent gets its own Workspace account and email address, so it works like another coworker, with context about teams, time zones, approvers and calendars. Actions are logged as the agent's, not a person's.
• Connections. Workspace, Microsoft 365, Slack, Jira, Confluence, Git, BigQuery, Databricks, Postgres, Snowflake and others, plus any Model Context Protocol server inside or outside the corporate network.
• Model choice. By default it routes each job to the model it judges best. Anthropic's Claude models were available at launch; Google says private and open models come later.
• Cost controls. Multi-model orchestration, smart routing and hard spending limits set in the Cloud Billing Console. When a cap is hit, the agent pauses, and resuming needs approval. Spend is tracked per project for chargeback.

Constellation Research reported that pricing is consumption-based with no separate SKU, that the exact consumption model for Workspace would be announced later, and that it expected general availability around the end of October or early November. As of 10 November I found no GA announcement.

The numbers

Why it matters

Three design decisions stand out, and none of them is about the model.

First, identity. Most enterprise agents today act with a delegated human token, so the audit log says "Priya edited the spreadsheet" when an agent did it. Giving the agent its own account and audit trail is what security teams have been asking for, and it is the problem I covered in non-human identity in 2026. It also means the agent's account needs the same lifecycle as an employee's: provisioning, least privilege, review and offboarding.

Second, MCP everywhere. Google committing to "any MCP server" means the connector work your team does for one agent platform is portable to another. That lowers switching costs at the tool layer, which is good for buyers.

Third, spend caps that actually stop work. A hard cap in the billing console that pauses the agent is the FinOps control most AI products still lack. Per-token prices fell around 98% since 2024, by Google's account, but volume grew faster, which is exactly the pattern that produces surprise invoices.

What sceptics say

Analysts were cautious about where lock-in moves. Nick Patience of The Futurum Group told CIO Dive that switching costs will sit in Google's layer, where agents accumulate memory, context and skills. Gartner's Joe Mariano questioned the line between Workspace and Gemini Enterprise: "I don't know if Google's been clear enough on that yet."

The adoption figures are Google's own and define "use" loosely. Nearly 90% of the Fortune 100 using Gemini Enterprise says nothing about how many employees use it daily or what it does for them. And an agent with its own mailbox, calendar and access to Git and Snowflake is a large blast radius if a prompt injection lands in one of the documents it reads. The governance features exist on paper; nobody outside the preview has tested them under pressure yet.

What it means for engineering and DevOps teams

• Treat agent accounts as service accounts. Decide who owns each coworker agent, what groups it joins, and how it is deprovisioned. Map its permissions to a role, not to whoever created it.
• Inventory your MCP servers now. If the agent can reach any MCP server inside the network, you need to know which ones exist, who wrote them and what they can do. The MCP ecosystem post covers registries and gateways for this.
• Set the cap before the pilot. Configure project-level spend limits and chargeback before the first team gets access. Pausing on a cap is only useful if someone is watching the approval queue.
• Check model routing against data rules. Automatic routing to Claude or later to third-party models may matter for data residency or contracts. Find out whether admins can restrict the model list per project.

FAQ

Is the Google Gemini agent available now?

Google launched it on 8 October 2026 in private preview for selected customers. Google said wide availability for Workspace customers on select Business and Enterprise plans is coming soon, without a date, and no general availability announcement had appeared by 10 November 2026.

Does the Gemini agent support MCP and non-Google models?

Yes. Google says it can connect to any Model Context Protocol server inside or outside the company network, and Anthropic's Claude models were available at launch alongside Gemini models. Open and private models are planned for later.

How do Gemini agent spending limits work?

Admins set hard limits in the Cloud Billing Console. When an agent reaches its cap it pauses, and continuing requires approval to exceed the budget. Spend is tracked by project so costs can be charged back to departments.

Related: Who Is This Agent? Non-Human Identity in 2026, MCP in 2026: 475M SDK Downloads a Month and FinOps for AI in 2026.