Google Cloud Launches Gemini Agent, One Universal Agent for Enterprise Work

At its Gemini at Work event, Google Cloud introduced the Gemini agent, a single cloud-hosted agent meant to handle questions, knowledge work, media creation and coding from one prompt box and one API. As MarkTechPost describes it, the pitch is delegation rather than chat: hand it an objective, and it plans the work, picks tools, connects to company systems and returns finished output, including jobs that keep running for hours or days after the laptop closes.

The piece that changes the conversation is the coworker agent. In Workspace, it gets its own account, with an email address, calendar, Drive and a directory entry. Colleagues can mention it in Chat or Docs, and its edits show up under its own name. That moves agents from a feature inside apps to something closer to a staff member on the org chart.

We think Google got the hard parts pointed in the right direction. Each agent gets an attested identity with least-privilege permissions, every action is logged to the agent rather than a person, traffic runs through an Agent Gateway that applies one policy across agents, and teams can set hard spend caps that pause a project's agent until someone resumes it. Those are the questions security and finance teams have been asking for a year: who is this thing, what can it touch, what did it do, and what does it cost. Answering them at the platform level is the useful kind of boring.

The model story is notable too. Google says the agent orchestrates across its own Gemini models and Anthropic's Claude models today, with other models planned, and routes each job to the best-fit, cheapest capable option. When a hyperscaler treats its own flagship as one option among several, the competition shifts from whose model is smartest to whose agent layer companies trust with their data and permissions.

Now the caution. MarkTechPost's own bottom line is that buyers have to evaluate this on customer anecdotes, not reproducible numbers. There are no benchmarks, no price and no general availability date yet. The headline customer figure, Bloomberg Media lifting SQL query accuracy by 63% by grounding data agents in Knowledge Catalog, is a real signal about the value of clean business definitions. It is not evidence about how the universal agent performs on messy, multi-day work.

The bigger thing I'm watching is what persistent memory and agent identity do at scale. An agent that builds a semantic map of documents and people, stores procedures it wrote for itself, and keeps an episodic record of everything it has done is powerful precisely because it accumulates. That same accumulation is a new concentration of sensitive context, and a new failure mode when it learns the wrong lesson and repeats it quietly under its own name. Version history that says the agent made the edit is accountability on paper; someone still has to read it.

Our read: this is one of the more complete enterprise agent designs we have seen, and the governance-first framing deserves credit. But a coworker that never sleeps, remembers everything, and holds standing access across Gmail, Salesforce, Snowflake and Slack is a big organizational change dressed as a product launch. Companies should pilot it like a new hire with limited access, not roll it out like a software update.

Google lists six design principles for the agent, including persistent cloud execution, temporary sub-agents with their own identities, and model choice per job. It connects to Slack, Jira, Salesforce, ServiceNow, BigQuery, Snowflake, desktop files and any Model Context Protocol server, and is reachable from web, mobile, desktop, CLI, Workspace, Microsoft 365, Slack or headless. MarkTechPost compares it with Microsoft 365 Copilot, OpenAI's ChatGPT Work and Amazon Quick.