The announcement in brief
Google is replacing a collection of assistants with one agent that accepts goals, not prompts. At its Gemini at Work 2026 event on 8 October, Google Cloud CEO Thomas Kurian introduced the Gemini agent, which he called “a single, universal agent for work”. It answers questions, does knowledge work, writes and runs code and hands parts of a job to sub-agents.
Two details matter most for engineering leaders. First, the agent picks a model per task from the Gemini family and Anthropic’s Claude models, with other private and open models promised later. Second, it connects to company systems through connectors and Model Context Protocol (MCP) servers, which is the same pattern teams use when they build custom AI agents today.
The agent is in private preview with selected customers. Google says it will be included in Gemini Enterprise at no extra charge, while pricing for persistent “coworker” agents will come closer to wider release.
What can the Gemini agent actually do?
The agent runs in Google’s cloud and keeps one memory and context across devices, so a task started in Gmail can continue from a laptop CLI. Google describes four memory types: session, semantic, procedural and episodic. Long jobs can run for days, and the agent creates sub-agents for parts of the work.
It works in three modes. As a personal assistant it knows the user’s calendar, team and documents. In delegation mode it suggests tasks it could take over. As a team “coworker” it receives its own Workspace account with an email address, calendar and Drive, and appears in the company directory.
For data teams Google showed PySpark generation, notebooks, model training and pipeline troubleshooting. For business users the agent answers reporting questions in plain language over BigQuery and Google’s Knowledge Catalog, which maps business terms such as “net margin” to the right metrics. Google says Bloomberg Media raised SQL query accuracy by 63% with Knowledge Catalog.
How does Google handle security and cost?
Identity and audit. Each agent has its own cryptographic identity and role-based permissions, and it reaches external systems through OAuth with least-privilege scopes. Actions are logged to an audit trail attributed to the agent, not to the person who started it.
Policy enforcement. Code runs in an Agent Sandbox with a network boundary. An Agent Gateway acts as a firewall for agents and applies company policy, for example “agents may not open need-to-know documents”, once for every agent.
Spend. Smart Routing sends simple steps to cheaper models. Project-level spend caps in the Cloud Billing console pause agents when a budget runs out, and costs can be charged back to departments. Kurian noted that per-token prices have fallen 98% since 2024 while enterprise AI volume has exploded, which is why budget control is now a product feature.
When is it available and what does it cost?
According to VentureBeat, the Gemini agent is in private preview, and Google has not given a date for general availability. It will be built into Gemini Enterprise with no separate setup or charge for existing customers. Pricing for persistent coworker agents, including their Workspace accounts and usage-based charges for long tasks, has not been published. TechCrunch reports that the launch starts with businesses and that a consumer version will follow later.
What it means for US & EU software teams
Generic agents are becoming a platform feature. Meeting scheduling, document drafting and simple reporting agents will now come bundled with the office suite. Budgets for custom agents should move to work that a general agent cannot do well: domain workflows, proprietary data and actions inside your own products.
MCP servers are now the integration contract. Google, Anthropic, Microsoft and OpenAI all support MCP. If your internal systems expose clean, permission-aware MCP tools, any of these agents can use them, and you avoid lock-in to one vendor’s agent.
Multi-model is the default, not an edge case. A Google product routing work to Claude confirms what many teams already do: pick models per task on cost and quality. Your evaluation sets, prompt versioning and fallback logic should not assume a single provider.
Governance must cover non-human identities. Agents with their own email accounts and audit trails are new identities in your environment. Under GDPR and the EU AI Act, teams in the EU need records of what data an agent can reach, who approved its permissions and how its outputs are reviewed. US teams under SOC 2 face the same questions about access reviews and logging.
What to do now
- Map your agent backlog and mark which use cases a bundled agent will cover and which need custom work.
- Inventory internal systems that agents will need, and plan MCP servers with scoped, read-first permissions.
- Set up model-agnostic evaluation so you can compare Gemini, Claude and other models on your own tasks.
- Define a policy for agent identities: who can create one, what it can access, how its logs are reviewed.
- Put budgets in place before pilots, with per-project caps and chargeback to the business owner.
Frequently asked questions
What is the Google Gemini agent?
The Gemini agent is a single enterprise AI agent that Google Cloud introduced at Gemini at Work 2026 on 8 October 2026. It takes objectives rather than step-by-step instructions, plans the work, delegates parts to sub-agents, writes and runs code and connects to company systems.
Which AI models does the Gemini agent use?
It routes each task to a suitable model from the Gemini family or from Anthropic’s Claude models, either automatically or by user choice. Google says support for other private and open models will follow.
Is the Gemini agent available now?
It is in private preview for selected customers. Google has promised wider availability soon but has not given a date. It will be included in Gemini Enterprise at no extra charge; pricing for persistent coworker agents has not been published.
How does the Gemini agent connect to company systems?
Through built-in connectors to tools such as Microsoft 365, Slack, Jira, Confluence, Git, Salesforce, ServiceNow, BigQuery, Databricks, Postgres and Snowflake, and through Model Context Protocol (MCP) servers plus enterprise tool and skill registries.
Does this make custom AI agents unnecessary?
No. Bundled agents will cover generic office work, but domain workflows, proprietary data and actions inside your own products still need custom development. The most useful investment now is clean, permission-aware MCP tools that any agent can use.
Sources
Google Cloud — Gemini at Work 2026: Introducing Gemini agent (8 October 2026)
TechCrunch — Google brings agentic AI to Gemini, starting with businesses (8 October 2026)
CNBC — Google Cloud introduces Gemini agent for work as AI race heats up (8 October 2026)
VentureBeat — Google Cloud unveils persistent Gemini Agents for long-running tasks (8 October 2026)