Google has turned Gemini into a worker with its own staff account. Thomas Kurian, the chief executive of Google Cloud, introduced the Gemini agent on 8 October at the company's Gemini at Work event in Mountain View, California, describing software that takes a goal rather than an instruction.
"You give it objectives, not just instructions," Kurian said. The agent plans the work, picks its own tools, connects to a company's systems and returns something finished.
Coworker agents are the part of the launch that changes how a business treats such software. Each receives its own Google Workspace account with an email address at agents.company.com, a calendar, a Drive and an entry in the company directory, and its edits appear in version history under its own name rather than a person's.
Google says tasks that run for hours or days keep going on its servers after the laptop closes, and are waiting when the person returns. The agent can also spawn temporary sub-agents, each with an identity of its own.
Staff reach it by typing @Gemini in Gmail, Docs, Sheets, Slides and Chat. In one example Google described, the agent was asked to arrange a meeting with a group whose members were never named. It worked out who they were from a chat space and an old email thread, checked calendars, and wrote to the external attendees itself.
What Google Did Not Announce
Google published no price for the Gemini agent, no general availability date and no waitlist. Its own post gives an availability stage only for the industry versions, which are in preview for financial services and legal work.
Several outlets describe a private preview for enterprise customers, and one says wider availability is coming for Workspace customers on selected Business and Enterprise plans. Google has not said either of those things itself.
What Google did publish is a way to cap the bill. Administrators can set a hard spending limit for each project in the billing console. The agent stops when a project hits its limit and waits for someone to approve more, and costs are tracked per project so departments can be charged separately.
Identity Is The Governance Argument
Google built the controls around four questions it put in writing: who the agent is, what it may do, where its actions can be seen, and what it must never touch. Identity answers the first two.
Every agent carries an identity the system can verify, governed like an employee's, with the narrowest permissions the job needs, and that identity is written into the logs and into any virtual machine spun up to run code for it. Actions are attributed to the agent rather than to the person who asked.
Each agent runs inside a sandbox, a walled-off space with its own network boundary, and all traffic between the agents and the outside passes through what Google calls Agent Gateway. That gateway is a network firewall enforcing company policy in real time, so a rule such as barring agents from documents marked need-to-know is written once and applied to all of them.
Google's materials set out no default requirement for a person to approve an agent before it sends, files or changes anything. They also publish no defence against prompt injection, the attack where hidden text on a page is read by a model as an instruction.
Models From A Competitor
The Gemini agent picks a model for each task and lets users override the choice. Google says it supports its own Gemini family today alongside Anthropic's Claude models, with other private and open models to follow.
Running a rival's models inside its own flagship product is unusual, and Kurian's materials present model choice as a feature rather than a concession. Google sells the agent layer rather than a single house model.
Connectors extend the same logic outward, reaching Confluence, Microsoft Office, Teams, Slack, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Snowflake and any MCP server. MCP, the last item on that list, is an open interface for plugging outside systems into AI tools that any vendor can implement.
Rivals Picked Different Shapes
OpenAI announced Workspace Agents in April 2026 as a successor to custom GPTs, built on Codex and able to run on a schedule and answer mentions in Slack. Sending an email or filing a ticket requires human approval by default there, a requirement Google's materials do not set out.
Microsoft went the other way with Agent 365, which governs agents rather than being one. It inventories agents built anywhere, gives each a Microsoft Entra identity and offers a switch to stop one mid-task. Anthropic sells Claude Managed Agents as a developer platform, charged at its usual rates for text processed plus $0.08 for each hour an agent is actively running.
Nick Patience, who leads AI platform research at The Futurum Group, put the trade plainly. "Model choice is open, but the memory, context and skills an agent builds up over time live in Google's layer," he told CIO Dive, adding that chief information officers "should go in with their eyes open."
Joe Mariano, a senior director analyst at Gartner, questioned where the product boundaries now sit. "Why would current clients of both Workspace and Gemini Enterprise need one of the other tools at this point?" he said. "I don't know if Google's been clear enough on that yet."
The Forecasts Point The Other Way
Gartner's published forecasts for this category read as a run of warnings. The firm expects more than 40% of agentic AI projects to be cancelled by the end of 2027 on cost, unclear value or weak risk controls. It also reckons only around 130 of the thousands of vendors calling themselves agentic are doing anything of the kind.
Gartner also expects 40% of enterprises to demote or shut down autonomous agents by 2027, because governance gaps only surface after something goes wrong in production. That forecast lands closest to the controls Google spent its launch describing.
Google noted that token prices have fallen 98% since 2024, and Gartner's costing runs the other way. The firm expects the cost of running each agent workflow to rise more than fivefold through 2028, because routing a task to a reasoning model costs providers at least five times what a basic chatbot exchange costs.
A Question Of Who Keeps The Context
Google reported nearly 500 cloud customers each processing more than a trillion tokens, and said close to 90% of the Fortune 100 use Gemini Enterprise. Those figures come from Google and cover its wider AI business rather than the agent announced this week.
The agent launched a day before this was written, so no independent data exists on whether companies will take it up. What the launch does establish is the shape of Google's bid, which is open on which model answers and closed on where the agent's memory, skills and audit trail live.
