Google Cloud's new work agent can route jobs to Claude and run for days

Google Cloud announced a Gemini agent at its Gemini at Work 2026 event on October 8th that can route jobs across the Gemini and Claude model families and run multi-step tasks for hours or days, with session, semantic, procedural and episodic memory kept in the cloud. The product is in private preview, with wider availability promised for customers on select Workspace Business and Enterprise plans, and Google says users can invoke it via @Gemini in Gmail, Docs, Sheets, Slides and Chat or through Slack, Microsoft 365 and the command line. Google frames the cross-model routing as improving output quality and lowering costs, but the announcement offers no comparative results for either claim.

Google Cloud's new work agent can route jobs to Claude and run for days Gemini agent will connect Workspace, third-party services and long-running sub-agents, but it is still in private preview. By RuntimeWire Staff https://runtimewire.com/author/runtimewire-staff · Published Primary source: 9to5Google https://9to5google.com/2026/10/08/gemini-agent-google-cloud/ Why it matters Google is trying to make its workplace AI handle work across Workspace and outside services, with model choice and persistent memory built in. The private-preview product still has to prove it can complete delegated work reliably. Google Cloud is pitching a new Gemini agent https://9to5google.com/2026/10/08/gemini-agent-google-cloud/?ref=runtimewire as a worker that can carry a task across apps, people and AI models, with workflows that may run for hours or days. Announced at Gemini at Work 2026 on October 8th, the product is in private preview; Google says wider availability for customers on select Workspace Business and Enterprise plans is coming soon. A user can give the agent an objective, Google says, and let it find the context, coordinate steps and use tools to pursue it. Google's bet is that the agent can manage work across an organization, including tasks that extend beyond a chat window. The work starts in Workspace Google's examples depend on access to the context already scattered through workplace software. In one, a manager emails for a project update as a slide deck; Workspace Intelligence identifies the request as suitable for delegation and offers a one-click handoff. In another, the agent is asked to arrange a meeting with a team's usual regional event leads. It identifies likely participants from a chat space and an earlier event thread, checks calendars and starts an email conversation, including with people outside the organization. Those examples build on Workspace Intelligence https://workspace.google.com/blog/product-announcements/introducing-workspace-intelligence?hl=en&ref=runtimewire , which Google introduced in April as a system for connecting information across Workspace apps, projects and collaborators. Gemini agent adds a proposed action layer: rather than only retrieving relevant material, it is meant to delegate and execute work using that context. Google says users can invoke it with @Gemini in Gmail, Docs, Sheets, Slides and Chat, or reach it through channels such as Slack, Microsoft 365 and the command line. Google also describes the agent as a potential team member, handling work for a group or performing a defined organizational role, such as a finance analyst. An API could let other services use it without giving it a dedicated interface. That flexibility could help Google put its agent in front of workers even when the task starts outside Workspace. One agent, several models Gemini agent can also use models beyond Google's own. Google says it selects the model it considers best suited to each job, using both the Gemini and Claude model families today, with private and open models planned later. Google frames that routing as a way to improve output quality and lower costs; the announcement offers no comparative results for either claim. The system can also send portions of a multi-step task to sub-agents, in parallel or sequence. Google's description includes work that continues for hours or days. It says the agent keeps session, semantic, procedural and episodic memory in the cloud, covering the task at hand, knowledge gathered from work, how procedures are performed and records of past actions. That architecture extends the direction Google Cloud laid out in April, when it described Gemini Enterprise as a platform for building, orchestrating and governing agents https://cloud.google.com/blog/products/ai-machine-learning/the-new-gemini-enterprise-one-platform-for-agent-development?hl=en&ref=runtimewire . Gemini agent is the user-facing promise of that platform strategy: one agent that can draw on Workspace context, delegate to other agents and choose among models. Google Cloud's broader proposition is that businesses can do this inside the services where their employees already work. The test is execution The product competes to make agents useful across an organization and across applications. Google's mix of Workspace access, outside services and third-party models is its product thesis. The hard part is whether an agent can act accurately and safely when it is reading workplace context, contacting colleagues and changing work across systems. The announcement's examples describe what Google expects the agent to do; the product remains in private preview. For customers assessing it, the consequential questions are how reliably it identifies the right people and documents, what permissions govern actions, and whether multi-step work costs less than having people coordinate it. Its value will be determined by the routine tasks it can complete without creating another round of checking and cleanup.