Google introduced a single Gemini Agent for enterprise work at its Gemini at Work event on October 8. The company describes an agent that can take an objective, use connected tools and company information, run code, and return work inside the applications employees already use. It also announced more data skills, industry versions, and controls for agent identity, access, model choice and spending. Google's launch post lays out that design; its current product page says several prominent capabilities remain in early access.

For an organization deciding what to deploy, the launch maps Google's intended system. The exact feature, edition and region still need confirmation before a team can treat it as a purchasable workflow.

One agent across work surfaces

Google says Gemini Agent can answer questions, create media, write and run code, and carry out delegated work through one agent and API. Its description includes temporary subagents for complex tasks and persistent coworker agents with their own identity, email, calendar and Drive presence. It says the agent can connect to tools such as Slack, Salesforce and Jira, as well as databases and Model Context Protocol servers. Reusable skills and a company registry are meant to encode how a team performs repeatable tasks.

The Workspace part brings the same agent into Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar. In Google's example, a coworker agent can join a team chat and produce a document under its own identity. The proposed benefit is continuity: a task assigned in a document or chat can use relevant company context and tools without starting over in a separate assistant. That is Google's product description, not a BIG CHANGE test of cross-application execution or permission boundaries.

The data announcement broadens the agent's proposed work. Google says engineers can ask for a data outcome and receive editable PySpark code and notebooks, while business users can request operational reports through BigQuery and Knowledge Catalog. Its announcement describes the catalog as the place to define business terms and locate datasets. A report produced this way would still need the organization's own checks on definitions, source data and generated queries before it is used in a decision.

Google is also packaging specialized skills and connectors for finance and legal work. The company labels those two industry offerings preview and says government, healthcare and retail versions are coming later. The announcement describes finance data sources and legal document permissions, but the preview label does not establish a general release or an independently validated compliance result.

Governance and cost are part of the product claim

The launch describes separate identities for agents, role-based permissions, action logs, isolated execution and an Agent Gateway for policy checks. These are consequential promises because an agent that can run code and act in external systems needs a way to identify who authorized an action and what it touched. Google's Agent Gateway setup documentation is more specific than the keynote: for Gemini Enterprise it documents outbound, agent-to-anywhere mode, scoped to the gateway's project and region. Administrators can select an audit-only dry run, which permits traffic while logging it, or enforce policies, which blocks requests without an explicit allow rule. Google recommends enforcement for production. Teams should verify the configured mode and log coverage rather than assume a gateway is blocking traffic.

For spending, Google says the agent can route work among Gemini and Anthropic Claude models and use Smart Routing to choose a model for each job. Its product page says third-party model choice is currently in early access. The cost controls have a more specific history than the October keynote alone suggests. Google's July 29 Spend Caps update described a public preview that sets a fixed monthly cap for one service within one project, with support listed for Gemini API, Agent Platform, Cloud Run and Cloud Run Functions. Google says the restriction takes effect within minutes of the threshold, so it is not an instantaneous cutoff. New on-demand charges stop for the capped service, while fixed commitments such as Provisioned Throughput continue to bill. An August 26 Gemini Enterprise billing update describes project-level agent API pauses, alerts and an overage option. The October keynote's project-cap language fits those later descriptions; it does not establish that every new Gemini Agent task, service or edition has identical coverage. Google has not published a separate price for the new agent in the material reviewed here. Buyers need their edition's cost terms and supported cap scope before budgeting against it.

Google's launch lists customer uses of Gemini Enterprise, including an insurer's thousands of custom agents and retailers' shopping assistants. Those accounts show that organizations are using parts of the broader platform. They do not establish that every newly announced Gemini Agent capability is generally available or that its security and cost controls have been independently tested. Performance and savings figures in the keynote are company-reported.

What buyers can verify today

The Gemini Enterprise product page offers a 30-day Plus trial and says to contact sales for broader adoption. It explicitly marks background delegation, mobile and desktop access, and third-party model choice as early access. Existing Gemini Enterprise release notes document individual features and regions separately; they should not be read as a blanket release notice for the new agent. Google has not supplied a single public matrix in the sources reviewed here that establishes the new agent's complete country coverage, edition entitlements, API availability and price.

For an evaluation, the useful unit is one permitted task in one actual edition and region. An IT team can ask Google which agent identity performs each action, where OAuth permissions are enforced, which tools pass through its configured gateway, which events appear in the audit trail and what happens when a project reaches its spend cap. A trial should then confirm those behaviors with the organization's own connected systems and policy settings. BIG CHANGE reviewed the public documentation and announcement; it did not run that trial.

The big change

What changed: Google presented Gemini as one work agent spanning chat, Workspace, data work and code, backed by reusable skills, connected tools and enterprise controls. Why it matters: A team could evaluate agent tasks alongside the permissions, logs and costs needed to operate them. What remains open: Several advertised capabilities are in early access. Spend caps have documented service, project and billing limits, and the sources reviewed do not give a complete new-agent availability or pricing matrix.

Sources & further reading