At its Gemini at Work 2026 event on October 8, Google Cloud announced Gemini agent, which it describes as a "single, general-purpose work agent." A user types a goal into one box, and the agent breaks it into steps, picks its own tools and gets the job done across apps including Gmail, Drive, Docs, Sheets, Slides and Calendar. It can also connect to Microsoft 365, Slack, Jira, Confluence, Salesforce and ServiceNow.
The enterprise edition is currently in private preview, and neither a general release date nor pricing has been announced. Google says it is rolling out to businesses first and consumers later, and there is no timeline for the consumer side.
A colleague with its own mailbox
The least chatbot-like part of this launch is the identity design.
Gemini agent can spin up multiple sub-agents as a task requires, running them in parallel or in sequence, and they can communicate and coordinate with one another. A second type, called a Coworker Agent, has a fixed role and a persistent identity: its own corporate mailbox, calendar, Drive space and directory entry. It can be added to Google Chat spaces, and colleagues can @-mention it to hand off work. It knows which team it belongs to, which time zone it works in and who its approver is.
Google stresses that these agents act under their own identity rather than borrowing the account of the employee who started the task, and audit records are logged under the agent's name. In a "task inbox," users can see its reasoning, how work was assigned to sub-agents, which skills were loaded, what code it wrote and how far along it is.
Google Cloud CEO Thomas Kurian framed it as agents that take objectives rather than just instructions:
"objectives, not just instructions"
Its capabilities fall into three parts: tools connected through MCP, Skills that store reusable procedures, and context. Memory is split into four kinds: session, semantic, procedural and episodic. The agent runs in the cloud, so tasks can continue for hours or even days, and it makes no difference if the employee closes their laptop.
Claude shows up in the model picker
By default the system picks a model automatically for each task, and users can also choose one manually. Gemini models and Anthropic's Claude family are supported for now. Google says it will expand to other closed and open models, as well as enterprises' private models. A single complex project can mix several models, and Google also mentioned a mechanism called Smart Routing, which classifies workloads to balance cost and quality.
How Claude is integrated and how it is billed is not explained in the InfoQ or TechCrunch reports, and we could not find it in Google's public materials either. Google is one of Anthropic's major shareholders, and Claude has long been available on Vertex AI. Putting it in the picker of Google's flagship workplace agent pushes that relationship one step further toward the front end.
On cost, administrators can set a hard AI spending cap per project in the Google Cloud Billing Console. The system tracks token usage and sandbox runtime charges, and when a cap is reached the corresponding agent is paused until an administrator manually resumes it. Permissions follow the principle of least privilege, Agent Sandbox provides an isolated execution environment, and Agent Gateway manages network ingress and egress. A single policy lets administrators bar all agents from files at a given confidentiality level.
Seen alongside Muse and Cowork
Three companies are building the same kind of product from different starting points.
Meta's Muse, launched in September, targets individual users in the US and Canada, with core features free and a paid subscription tier. On September 29 it added Muse for Small Business, which connects to Asana, Notion, Shopify, Slack and Stripe. It is powered by the Muse Spark model, each user's agent runs in its own cloud virtual machine, and outward-facing actions such as sending email or making payments require user confirmation.
Anthropic's Claude Cowork takes a route built around the employee's personal desktop plus enterprise controls, and it recently moved its whole runtime environment to the cloud.
Google's version puts the weight on enterprise identity, permissions and unified policy, plus multiple models. For companies already deep in Workspace, the migration cost is lowest; for companies centered on Microsoft 365, it can connect, but that is not its home turf. For teams in China, using it requires access to Google Workspace, and that condition alone rules out most of them.
Google's own scale figures are that Gemini has more than 1 billion monthly active users and that nearly 90% of the Fortune 100 use Gemini Enterprise at work. Early trial participants include the sports brand On, Shopify and PayPal. Finance and legal editions are in preview.
Sundar Pichai also acknowledged at the event that agents still have harder problems to crack in security, scale and performance. When the preview ends, pricing and Claude's billing arrangements are expected to be revealed together, and that is when we will see whether enterprises are actually willing to pay for "one more colleague."
Sources: TechCrunch, InfoQ, CocoLoop, Google Cloud public event information; figures checked include Gemini monthly active users, the Fortune 100 usage share, the private preview status and the range of supported models.