Google Cloud has launched Gemini Agent, a single universal agent for work, announced in a blog post signed by Google Cloud CEO Thomas Kurian at the company's Gemini at Work 2026 event on October 8, 2026. The agent can research, draft email, build presentations, analyze data, write and run code, call tools across applications, plan its own tasks, and assemble temporary teams of sub-agents for complex work.
The product is available now in private preview to a select group of customers. Google expects general availability by the end of October 2026, starting in North America, according to a report citing the company; a Google Cloud spokesperson separately told VentureBeat that wider availability was coming soon without giving a specific date.
An AI colleague with a real enterprise identity
Under the Gemini Agent framework, Google introduced what it calls Coworker Agents — persistent digital workers that a company can create by describing a role. Each coworker agent receives its own Workspace account, including an email address, calendar, Drive storage, and a presence in the company directory, according to Google's announcement.
Teammates can then treat it much like a human colleague: invite it into Google Chat spaces, @mention it in documents, or email it assignments. Because the agent acts under its own identity, its actions are logged under its own name rather than a person's, which Google says makes its work easier to audit and trace.
Google has not clarified whether every Gemini agent receives a full Workspace account or whether that is reserved for coworker configurations, nor whether additional Workspace licenses are required.
Four kinds of memory
Gemini Agent is built around four memory types, described in Google's announcement and in reporting from QbitAI and VentureBeat:
- Session Memory — holds the context of a task in progress, even one that runs for days.
- Semantic Memory — accumulates structured knowledge from documents, conversations, and other agents, such as who owns which project or what an internal term means.
- Procedural Memory — records how a job should be done, including data sources, workflows, and delivery formats; the agent can also write its own reusable "Skills."
- Episodic Memory — stores what has been completed in the past and how.
Google's framing is that the agent should learn like a new employee: the first time it builds a weekly report, a user may need to explain the structure and data sources; by the third time, it should reuse what it learned.
A model layer that can change
Gemini Agent does not have to run on Gemini. Google says the agent routes each job to the model that fits best, drawing on its own Gemini models and Anthropic's Claude models today, with OpenAI and open-weight models to follow.
A companion Smart Routing capability selects models per workload based on performance, speed, and cost — cheaper models for simple jobs, stronger models for complex ones — and multiple models can be combined within a single project. Administrators can restrict which models are available to agents, Google confirmed to VentureBeat.
The architecture also decouples the agent layer from the model layer: the underlying models can change while the agent's accumulated context, skills, and enterprise data persist, so companies would not have to rebuild established workflows if model rankings shift.
That flexibility comes against a mixed benchmark record for Google's newest flagship. In the comparison tables Google published with Gemini 4 Argon on September 30, 2026, Argon leads Claude Opus 5.5 on DeepSWE v1.1 (77.9% versus 74.2%) but trails it on Terminal-bench 4.0 (57.4% versus 66.4%) and FrontierSWE v2 (55.0% versus 62.3%). Those are Google's own published figures; independent verification had not been published at the time of the agent launch.
Pricing and governance
Once broadly released, the agent will be built directly into Gemini Enterprise with no separate setup or dedicated website, and will be included at no additional charge for customers where Gemini Enterprise is available, including existing Google Workspace business and enterprise subscriptions, VentureBeat reported.
On governance, Google says every agent receives a cryptographically attested identity with least-privilege permissions, executes complex tasks inside an isolated Agent Sandbox, and passes traffic through an Agent Gateway that enforces company policies in real time. Google also announced project-level spend caps that pause agent activity when a budget is exhausted.
A crowded field
The launch lands in a competitive window. Microsoft introduced an updated Copilot on September 25; OpenAI announced its always-on Dots agents on September 29; Meta has pushed its personal agent Muse; and Anthropic announced native Claude integration with Google Docs, Sheets, and Slides on October 6 — two days before Google's event.
The rivals differ in emphasis: Meta's Muse leans toward personal errands such as shopping and travel, OpenAI's Dots centers on continuously working on a user's behalf across connected apps, and Google's pitch is the deepest integration into an enterprise's existing organization — its directory, permissions, mail, calendars, and documents.
What is confirmed, and what is not
Confirmed: Gemini Agent exists, is in private preview, supports persistent Coworker Agents with Workspace identities, uses four memory types, and can already route tasks between Gemini and Claude models with more vendors planned.
Unknown: The exact general-availability date beyond the reported end-of-October 2026 target for North America; whether coworker accounts require extra Workspace licenses; whether data is transmitted to external model providers such as Anthropic when third-party models are used, and on what retention terms; and whether administrators can toggle individual services per agent.
Still awaited: Independent benchmarks on how reliably the agent completes long, cross-application assignments — Google has not supplied any — plus further pricing details promised closer to wider release.