One agent instead of a product line
Google Cloud introduced the Gemini agent on 8 October, calling it “a single, universal agent for work”. The post adapts a keynote by Thomas Kurian, the chief executive of Google Cloud, at the company’s Gemini at Work conference.
The framing is a consolidation. Where Google has sold assistants, platforms and copilots as separate things, it now describes one prompt box and one API that answers questions, does knowledge work, generates media and writes and runs code. Users give it objectives rather than steps, and Google says it returns finished work inside the documents, inboxes and developer tools people already use.
It runs on web, iOS, Android, Windows and Mac, from the command line, and inside Google Workspace, Microsoft 365 and Slack. It can also be embedded in third-party applications or run with no interface at all.
It will run your work on Claude
The detail worth pausing on is the model router. Google says the agent orchestrates across its own Gemini models and Anthropic’s Claude models today, with other private and open models planned, and that each job runs on whichever model it judges best suited. A feature called Smart Routing makes that choice automatically, aiming for the lowest cost at acceptable performance.
That is a cloud provider telling enterprise buyers it will spend their money with a direct competitor when the competitor’s model is the better fit. Google’s own model line-up — Argon for frontier reasoning, Flash for volume, Omni for generative media, Gemma for the edge — is presented as one input to that decision rather than the product itself.
Agents with their own email addresses
The agent creates temporary, job-specific sub-agents with their own identities and coordinates them in parallel or in sequence. It also introduces what Google calls a coworker agent: a persistent, role-based agent with its own identity, its own storage, and access only to the context its team gives it. In Workspace that agent gets its own account and its own address on a dedicated domain, can be added to chats or tagged in document comments, and appears in version history under its own name.
Google describes four kinds of memory behind this — session, semantic, procedural and episodic. Procedural memory covers how work gets done, including skills the agent has written for itself.
A firewall for agents
The governance section is organised around four questions: who the agent is, what it may do, what it did, and what it must never touch. Each agent gets a cryptographically attested identity that is stamped into logs and into any virtual machine it starts. Agents run inside an Agent Sandbox with their own network boundary, and all their traffic passes through Agent Gateway, which Google calls an AI network firewall and which can enforce rules such as barring agents from documents classified need-to-know.
Cost gets a similar treatment: hard per-project spending caps set in Cloud Billing, which pause that project’s agent when hit and resume it with one click.
The numbers are Google’s own
Google says nearly 500 customers each processed more than a trillion tokens in the past year, that nearly 90% of the Fortune 100 use Gemini Enterprise, and that nearly 80% of its cloud customers use its AI products. It claims its TPU 8i delivers 80% better price-performance than the previous generation. None of those figures is independently verified, and the customer results it cites — Bradesco cutting a document review from an hour to five minutes, Ulta Beauty tripling digital conversion — are reported by the customers.
Industry versions are further along in some sectors than others: Financial Services and Legal are in preview, with Government, Healthcare and Retail described as coming soon.
What to watch
The post does not give a general availability date or any pricing for the Gemini agent, which makes the announcement a direction rather than a shipping product. The thing to watch is the router: if Google’s own tooling keeps sending frontier work to Claude, the question of which lab wins enterprise AI starts to look less important than who owns the orchestration layer above them.
