Google Cloud launches Gemini Agent for office work, supports calling Claude

Google Cloud published a long-form post titled 'Welcome to Gemini at Work 2026: Introducing the Gemini agent,' signed by Google Cloud CEO Thomas Kurian, officially launching Gemini Agent. According to Google's vision, this is a general-purpose office Agent covering various work tasks and capable of long-term operation—it can look up information, write emails, make slides, analyze data, and write and run code.

But the most unusual point in this release is that Google explicitly states the underlying model of Gemini Agent can be selected separately: at present, it can dynamically choose between its own Gemini series and Anthropic's Claude series, with plans to support more closed-source and open-source models in the future. Google calls this capability multi-model orchestration and pairs it with Smart Routing for automatic routing, choosing between effectiveness, speed, and cost based on the task.

A company voluntarily listing a competitor's model as an optional foundation for its own Agent indicates it judges the focus of competition is no longer the model itself. This is the inference of this article, based on another statement Google emphasizes at the same time: the underlying model can change, but the context, Skills, and enterprise data accumulated by the Agent can be retained, and enterprises adjusting the underlying model need not tear down established workflows. If context and workflows are the assets, then the model becomes a replaceable part.

Google's working principle for Gemini Agent is 'You give it objectives, not instructions'—that is, users only need to state the goal rather than step-by-step instructions. The officially described core capabilities include a unified entry point, persistent cloud operation, cross-platform invocation, multi-Agent collaboration, event triggers, and scheduled tasks.

It should be noted that the above are all Google's official descriptions, i.e., vendor self-statements, with no third-party independent verification. The relay article also does not give a specific release timeline, pricing, or availability regions for Gemini Agent, nor does it provide any benchmark scores or performance comparison data. Therefore, 'what it can do' can currently only be understood according to the official line, and actual usability and stability remain to be seen.

Under the Gemini Agent framework, Google proposes the Coworker Agent: allowing enterprises to create a long-existing identity for Gemini with clear work responsibilities, configured with an independent Google Workspace account, enterprise email, calendar, and Google Drive storage space.

This identity is not decorative. According to the official description, the AI coworker can appear in the company directory, and team members can invite it to Google Chat group chats or @ it in documents just like a regular colleague; it performs operations under its own identity and leaves its own records, making it easy for enterprises to manage and track what it actually did.

Transforming the Agent from a 'tool' into an 'organizational member' changes the object of governance. Tools only need permissions; members need identity, auditing, and accountability. The relay article mentions that Google puts model selection, enterprise data connection, security auditing, and cost control into this system, but the permission and audit risks after enterprises adopt an independent-identity AI coworker remain unclear—this point is still a blank in the available information.

According to the official description, Gemini Agent has four types of memory. Session memory is used to preserve the context of the current task; even if the task lasts for days, it can remember completed steps and pending work. Semantic memory allows it to accumulate knowledge from reading documents, communicating with people, and collaborating with other Agents, organizing it into a structured knowledge base. Procedural memory records how a job should be done; it can even write its own Skills, saving methods learned from work for use in subsequent tasks. Episodic memory is used to record what work has been completed in the past and the corresponding execution experiences.

The goal of combining these four types of memory is straightforward: reduce the need for users to repeat themselves. Google uses a metaphor, saying Gemini will, like a new employee, gradually become familiar with you, your tools, and your team before formally starting work.

But the stronger the memory, the sharper the data boundary issue. The effectiveness of the four memory mechanisms in real enterprise environments and the way data is isolated are unclear in the available information; the actual trade-offs of multi-model orchestration between effectiveness, speed, and cost are equally unknown. These are not questions that a launch copy can answer.

The timing of Google's entry is delicate: on September 29, OpenAI had just released Dots, capable of working 24/7, and Meta had previously launched its personal AI Agent Muse.

According to the relay article's description, OpenAI's Dots has an independent cloud computer, supports 24/7 operation, can remember users' long-term goals and work habits, and can connect to over 4,000 applications, but it emphasizes continuous work around the individual user, and its professional Dots for enterprise roles is mainly in early enterprise pilot stages. Meta's Muse leans more toward personal life scenarios. In contrast, Google's focus this time is on letting the Agent enter the enterprise's existing organizational and office systems, using business context in applications like Gmail, Docs, Sheets, and Calendar to understand what employees are doing.

It should be noted that these descriptions of Dots and Muse also come from the relay article, without first-party sources attached, and should be regarded as source claims rather than confirmed facts. The final direction of the three-way competitive landscape cannot be judged from the available information.