What Google announced
Google Cloud CEO Thomas Kurian introduced the Gemini agent on October 8, 2026, in a blog post tied to the Gemini at Work 2026 event. Google describes a single agent and a single API that handles questions, knowledge work, media creation and coding.
The pitch is delegation by outcome. Constellation Research quotes Kurian saying users give the agent "objectives, not instructions." The agent then plans the work, picks skills and tools, connects to company systems and returns finished output.
That is the real difference from a chatbot. The claim is execution across connected systems, persistence and orchestration. Plain conversation is not the selling point.
Where it runs, including Windows and Microsoft 365
Google says the agent can be reached from:
- web and iOS/Android mobile devices
- Windows and Mac desktops
- the command line
- Google Workspace, Microsoft 365 and Slack
It can also be embedded in third-party applications. It can run as a "headless" agent with no dedicated interface. VentureBeat reports that for existing customers it will work through Gemini Enterprise and familiar workplace interfaces, including Microsoft 365, rather than requiring a separate app.
Google says execution happens in the cloud. Memory, context and personalization therefore follow the user across devices. Work that takes hours or days keeps running after the laptop closes.
Connectors and tools
Google lists these connection targets:
- Collaboration: Confluence, Microsoft Office, Teams, Slack and Workspace.
- Development: Git and Jira.
- Enterprise platforms: Salesforce and ServiceNow.
- Databases: BigQuery, Databricks, Postgres and Snowflake.
- Local files: files on a user's desktop.
The agent can also talk to Model Context Protocol (MCP) servers inside or outside the corporate network. An enterprise tools registry lets teams publish internal tools for colleagues.
Google has not detailed setup requirements, licensing dependencies or whether each connector offers the same capabilities. Google's earlier Gemini Enterprise materials mention pre-built connectors for Outlook, SharePoint Online, OneDrive and Microsoft Teams. Whether the new agent uses those same connectors is not spelled out in the announcement.
Sub-agents, skills and memory
Google says the agent can act as a personal assistant, a team member or a role-based worker, such as a finance analyst. It can spin up temporary sub-agents, each with its own identity, for parallel or sequential steps. It can also act as a "coworker agent" with a persistent role. Coworker agents get their own address in the form @agents.company.com and their own storage. Google says they see only the context shared with them.
Skills are reusable modular instructions or workflows. Organizations, departments and individuals can publish them to registries. Google describes four memory types:
- Session memory for the current task.
- Semantic memory, a structured knowledge base built from documents and interactions.
- Procedural memory for how work gets done, including skills.
- Episodic memory of past actions.
Google does not say how long any of these memories are kept. It also does not say how users or admins can inspect or delete them, or what the default data-governance settings are. These are open questions, not confirmed flaws.
Model routing
The agent is separate from the model underneath it. Google says it currently routes among its own Gemini models and Anthropic's Claude models. It plans to add other private and open models later. SiliconANGLE reports that simple tasks might go to a small, fast Gemini Flash model, while long-horizon work could go to a flagship model.
Google says the early testers include On, Shopify and PayPal. Constellation Research also names those three. Google supplies no benchmarks or pricing for the routing feature.
Governance and cost controls
For IT administrators, this section matters most. Google says:
- Identity: each agent has its own cryptographically attested identity and least-privilege permissions.
- Permissions: access is role-based and fine-grained. When the agent connects to external systems, identity is mapped and propagated through standards such as OAuth.
- Auditing: every action is logged and attributed to the agent, not to a person. Admins can watch the logs through observability tools.
- Policy: agents run in an Agent Sandbox with a network boundary. All traffic in, out and between agents passes through Agent Gateway, which Google calls an AI network firewall. An example policy is that agents may not open documents classified "Need to Know."
- Spend: Smart Routing picks the model per workload. Real-time spend caps are set per project in the Cloud Billing Console. When a cap is hit, that project's agent pauses, and an admin can resume it from the console.
These are Google's descriptions. The announcement offers no independent validation, deployment prerequisites or availability tiers.
Availability: still early
Despite the broad vision, the rollout looks early. 9to5Google reports the agent will see wide availability soon for Workspace customers on select Business and Enterprise plans, and that it is currently in private preview. VentureBeat adds that Google has not said exactly how customers activate the universal agent or when all its capabilities will arrive.
Google also says its financial-services and legal specializations are in preview. Government, healthcare and retail are listed as "coming soon." Nothing in the announcement points to general availability for the whole feature set, and no pricing has been published for it.
Data and analytics extras
Google bundled several data features with the launch:
- Engineering: the agent can generate PySpark code, provide editable notebooks, train models and troubleshoot pipelines.
- Reporting: business users can ask for reports that become saved queries. Google says these rerun on demand without token costs.
- Knowledge Catalog: maps business definitions so agents share them.
- Smart Storage: enriches unstructured files in place.
- Borderless Lakehouse: queries Amazon S3 and Azure Data Lake without variable egress fees. It federates Iceberg tables across Databricks Unity, Snowflake Polaris and AWS Glue.
Google cites Bloomberg Media as lifting SQL query accuracy by 63% during early development after grounding its data agents in the Knowledge Catalog. That is a company-reported example, not an independent test.
Customer claims
Google's post lists many Gemini Enterprise deployments. Here are three, all as Google presents them:
- Orange Spain has deployed more than 1,000 custom agents.
- SOMPO built over 10,000 agents across 34,000 employees.
- Bradesco reports document review dropping from one hour to five minutes, with a 60% cut in risk inconsistencies.
These come from the broader Gemini Enterprise platform. They are not benchmarks of the new agent.
A caution on the hype
The Tech Buzz piece pitches the launch as an enterprise automation play. The sources I could verify do not back every example in that coverage, such as specific CRM updates or approval routing. They also do not support a head-to-head claim against Microsoft Copilot. What Google has documented is a cross-system agent ambition with Microsoft 365 and Teams compatibility. Nobody has shown it is reliable for mission-critical processes or that it matches Copilot.
Also note that Google's detailed inline examples, such as meeting coordination and coworker agents with their own accounts, are Workspace scenarios. Nothing in the announcement shows equivalent behavior inside Microsoft 365.
What Microsoft-based IT teams should do
Think of this as a pilot checklist, not a purchase order.
- Map the connectors. Find out which Microsoft services the agent would reach, and under which tenant permissions and consent grants.
- Check identity propagation. Confirm how agent identities appear in Entra sign-in and audit logs when OAuth is in play.
- Test audit visibility. Verify that agent actions can be traced end to end across both vendors' logging.
- Test policy enforcement. Try cases such as sensitivity-labeled documents to see whether the Agent Gateway policy model respects your classification scheme.
- Set spend caps first. Per-project caps exist, so use them from day one.
- Ask about memory. Press for retention periods and deletion controls before sensitive data goes in.
Bottom line
Google is positioning one agent that reaches across Workspace, Microsoft 365, Slack and a long list of enterprise systems. The governance story, with its agent identities, audit trails, gateway policies and spend caps, is the part worth scrutinizing. Right now it is a private-preview announcement backed mainly by Google's own claims. Microsoft shops should watch it and test it carefully, and should not assume it can replace or match their existing Copilot setup.
Update: Additional details (October 8, 2026)
VentureBeat reports that persistent “coworker” agents can receive a Workspace account with their own Gmail address, Calendar, Drive storage and directory presence—not merely an @agents.company.com address and storage. Staff could assign work by mentioning an agent in Google Chat or asking it to edit a document. Google describes this as one configuration rather than a requirement for every Gemini agent.
The report also says agent activity is logged when an agent creates isolated code-execution environments. It adds that Google has not specified whether administrators can independently enable or disable the Gmail, Calendar, Drive and directory components, nor how their provisioning and licensing will work.
Update: Report adds October rollout target and consumption-billing details (October 10, 2026)
Google reportedly expects the Gemini agent to reach general availability by the end of October 2026, starting in North America, according to Futurum’s October 9 analysis. That adds a specific rollout target to the earlier availability reporting, but remains an expectation—not confirmation that the private preview has ended or that every announced capability will ship together.
Futurum also reports that Google plans no additional agent-seat charges where Gemini Enterprise is available, while advanced, long-running work would draw on pooled quotas or consumption billing. This adds pricing detail absent from the initial coverage, rather than establishing that all agent work is free. Microsoft-centric IT teams should distinguish access entitlements from usage costs and obtain applicable quota and billing terms before budgeting a deployment.
According to Futurum, administrators will control which models their organization permits, while users can select a model for each task or leave routing automatic. The outlet also reports that per-user spending caps are planned alongside the already announced per-project caps, and that OpenAI models are intended to join the routing roster. Those remain reported future additions; administrators should not assume either is available in the current preview.
References
- Google Cloud unveils persistent Gemini Agents for long-running tasks, and they get their own Gmail, Calendar, and Drive storage VentureBeat · 2026-10-08T12:00:00+00:00
- Google Cloud Launches Gemini Agent for Enterprise Automation - The Tech Buzz The Tech Buzz · 2026-10-08T12:37:15.902000+00:00
- Google Gemini Agent Puts the Agent Ahead of the Model The Futurum Group · 2026-10-09T16:21:41