On October 8, 2026, at its Gemini at Work event, Google Cloud introduced a Gemini agent for Gemini Enterprise customers. Chief executive Thomas Kurian summarized the pitch as “you give it objectives, not just instructions.” The agent plans multi-step work that can run for hours or days in the cloud, spans Google Workspace, Microsoft 365 and Slack, and keeps context across desktop, mobile and web. It is in private preview for selected customers, and Google has not publicly confirmed a release date, although executives reportedly told analysts to expect general availability around late October or early November.
The headline feature is less about intelligence than about identity. Businesses can create persistent coworker agents that receive their own account in the Workspace directory.
An Agent With a Badge
According to launch coverage, each coworker agent gets an address ending in @agents.company.com, a calendar, Drive storage and a company-directory entry. Colleagues can mention it in Chat or tag it in a Doc comment, and its edits appear under its own name in version history. Google says the agent sees only what has been shared with it. Each agent carries a cryptographically attested identity, with permissions that are role-based and admin-approved, and its actions are written to an audit trail attributed to the agent rather than to the person who asked.
Behind the scenes, the system routes work across Gemini models and Anthropic’s Claude models. Administrators can restrict which models agents may use, and a feature called Smart Routing picks among the permitted ones by task and cost. Google says prompts, documents and chat histories stay inside Gemini Enterprise when Claude is used, but it has not said whether data goes to Anthropic, how long it is kept or where. Industry versions for financial services and legal work are in preview, with the finance version reportedly including more than 50 skills and data from FactSet and LSEG.
What Google Has Not Yet Explained
The design answers the first questions an auditor would ask: who is the agent, what may it do, and what did it do. It leaves gaps. Public material does not explain whether audit logs carry a delegation chain, meaning who requested or approved a given action. It does not say whether the identities of subagents are revoked when a job ends, or whether a subagent’s permissions can exceed its parent’s. One analysis also raises a mundane but real question: who owns an agent’s account when the employee who created it leaves the company?
Cost control is similarly partial. Google says hard spend caps, first announced in August, now apply to the agent, and that when a project hits its limit in the Cloud Billing Console its agent pauses and can be resumed with one click, with per-project tracking for charging costs back to departments. But Google’s own spend-cap documentation, updated October 7, describes limits that matter: caps apply to one eligible service in one project on a monthly basis, are based on estimated list prices, and are not instant, so delayed overages can still be billed. In-flight requests finish and bill, and subscriptions, including Gemini Enterprise subscriptions, are out of scope. It is not stated whether the agent’s cap uses this exact mechanism.
Pricing is unsettled too. Google says there is no additional charge for Gemini Enterprise customers, but charges for persistent coworkers or extended autonomous tasks are unspecified, and pricing is to be shared closer to the wider release. There are also no independent benchmarks of how reliably the agent finishes long, multi-application tasks. Analysts quoted in coverage were measured: one noted the launch creates no new category, since OpenAI, Anthropic, Microsoft and Meta already offer multi-agent capabilities, and another questioned whether security leaders can trust such an agent with mission-critical processes over time.
Why the Direction Still Matters
Skepticism about maturity should not hide the structural shift. Google is treating an agent as a managed member of the organization rather than a feature inside an app: it appears in the directory, takes comments, is billed to a cost center and can be paused by finance. That is the same bet other vendors are making in different words, and it implies that the central enterprise question for the next year is governance rather than raw model quality. The UK Information Commissioner’s Office, notably, opened a call for evidence on agent risks the same day, with responses due November 20.
There is a competitive angle as well. By routing across Gemini and Claude and connecting to Microsoft 365 and Slack, Google is signaling that it intends to be the orchestration layer even where a customer’s data lives elsewhere. That puts pressure on smaller agent startups whose pitch is cross-app automation, because the largest platforms are now bundling it.
What This Means for Philippine Founders
For Philippine startups selling software to businesses, the lesson is to design for the questions Google is still answering. Enterprise buyers, including the multinationals that outsource work to the Philippines, will ask who an agent acts as, what it can touch, how its actions are logged and how its spending is capped. A product that can show agent-level audit logs, scoped permissions and hard usage limits will be easier to approve than one that cannot, regardless of which model sits underneath.
There is also a services opportunity. If coworker agents arrive in Workspace tenants, companies will need people to configure sharing rules, review agent permissions and write policies for who owns an agent account. Philippine IT-services and outsourcing firms that already administer client workspaces could package agent governance as an offering. Founders should also note the data-residency ambiguity around Claude routing: until Google clarifies where data goes, regulated clients such as banks will likely ask for written answers, and a startup that has them ready will move faster.
Share this article