A universal work agent is easiest to buy as a simple interface and hardest to evaluate as a system of authority. When an assistant moves from answering a question to editing a record, creating a presentation and continuing work in the cloud, the important question changes: what can this identity do when the employee stops watching? Google's October 8 Gemini Agent announcement makes that question concrete.
Google describes one persistent agent across workplace applications and devices, with shared memory, skills and permissions. It can orchestrate Gemini and Claude models; more private and open models are planned. This is an orchestration claim, not evidence that every underlying model is interchangeable without changing quality, cost or data-processing requirements.
Define one outcome before connecting everything
Start with a bounded task, such as preparing an internal project update from approved documents. Specify the required sources, calculations, output format, completion criteria and reviewer. Keep external sending and consequential record changes outside the pilot. A useful baseline includes the existing human process, not merely another model's response speed.
Google lists Workspace, Microsoft 365, Slack and enterprise data connectors as part of the agent's reach. These surfaces do not remove integration work. Check identity mapping, freshness, read versus write scopes and connector failures individually. A common prompt box can hide several authorization systems without making their rules consistent.
Follow delegated authority through every worker
Google says coworker agents can have their own email address, calendar, storage and directory entry. Temporary sub-agents also receive identities. That could make actions more attributable, but an identity is not itself a safe permission policy. Verify which grants a child receives, how long they last, and whether logs distinguish its actions from the parent's.
The announced sandbox and Agent Gateway provide execution and policy boundaries. Test those boundaries with revoked access, sensitive classifications, compromised tools and hostile instructions inside ordinary documents. Permission to read a document must not mean permission to obey it. NIST's identity-and-authority discussion is useful context for least privilege, delegation and auditing, rather than a certification of this product.
Treat memory as governed data
Persistent context can prevent employees from restating a project when work moves between applications. It can also preserve an incorrect assumption after its original conversation. Identify who can inspect, correct, restrict and delete personal, project and organizational memory. Test a role change and a withdrawn source, including derived documents, rather than checking only the original connector permission.
Skills need owners, revision histories and withdrawal procedures too. A stale credit-memo template can spread the same error across many successful-looking tasks. Google's DBS example of deterministic specialized-agent chains with people involved suggests a useful distinction: one broad interface can still sit above narrow, explicitly controlled workflows.
Measure finished work and the cost of correction
Google reports substantial adoption and customer results, including reduced development time, administrative hours saved and conversion gains. These are company-reported, workflow-specific examples, not a reproducible success rate for unrestricted delegation. Catalog shopping assistance and open-ended employee work have different failure surfaces.
Score the entire pilot: source preparation, task completion, factual corrections, reviewer time, policy violations, unintended actions and failure recovery. Project spending limits can pause work, but they do not prove value. Include model routing, sub-agent usage and human rework in total cost, and verify an interrupted task cannot continue spending through another identity.
The durable purchasing question is whether the control layer remains understandable as usage grows. Before expanding beyond the pilot, test exportability of skills, memory, evaluations and policies. Supporting another provider's model does not automatically make enterprise context portable. Expand authority only when measured completion and policy compliance justify it, not because a universal interface makes delegation feel familiar.