Google announced its Gemini agent for work on October 8, 2026, presenting enterprise buyers and platform teams with a single workplace agent for questions, document work, media creation and code. The proposed benefit is shared business context across tasks that otherwise require separate tools and repeated handoffs.
Our assessment: pilot a supported, reviewable workflow before standardizing. The launch broadens the integrated option, but access, data handling and billing must be checked for the specific features your team needs. The published material does not establish that this approach is cheaper or more reliable than a custom system.
What is announced, and what can you actually use?
Google describes coworker agents with their own Workspace accounts, separate from temporary, task-specific subagents. Coworkers act under their own identities and receive context through team sharing; Google also describes cloud execution that continues after a user closes their laptop. These are announced product capabilities, not proof that every edition or tenant has them enabled.
The current product FAQ explicitly marks multistep background delegation, mobile and desktop access, and third-party model choice as early access. It offers a 30-day Gemini Enterprise Plus trial, but that trial should not be treated as confirmation of access to every advertised capability.
Model availability needs particular care: the launch announcement says routing across Gemini and Claude is available today, while the FAQ says Gemini models are available now and additional models are coming. The reviewed pages do not reconcile those scopes. Confirm the required model in your tenant before making it a production dependency.
Nor is every infrastructure component new: Agent Gateway, Agent Registry and Agent Observability were already generally available in the June 18, 2026 platform release notes. Their release status does not establish general availability for all of the newly announced workplace experiences.
Choose the architecture around the task
The following is a decision framework, not a tested product ranking. First establish whether the job needs an agent at all: a fixed sequence of known lookups and approvals may still suit a conventional workflow.
Your requirement | Starting choice | Evidence needed before expansion |
|---|---|---|
Supported document or data work; Google-managed execution is acceptable. | A bounded integrated pilot. | Enabled features in the target edition and region, approved connector permissions, and a checkable output. |
A critical step requires early access, a particular external model, or unsupported local processing. | Retain the existing production path; isolate the experiment. | Explicit access and feature-specific commitments, rather than a general launch promise. |
A narrow dataset, fixed business rules, custom recovery, or a redeployment requirement. | A curated component or custom/hybrid workflow. | Enforceable dataset limits, defined recovery behavior, and a documented state-export or rebuild plan. |
Custom does not necessarily mean leaving Google’s ecosystem. Its open-source Agent Development Kit (ADK) supports multiple model providers and deployment on your own infrastructure or Google Cloud. That offers an implementation option, not automatic portability of the new workplace agent’s memory, policies or task state. Those export guarantees were not established in the reviewed sources.
Separate data discovery from controlled reporting
Google’s Data Cloud connector documentation makes a useful distinction. Ingestion copies source data into a Gemini Enterprise data store. Federated querying accesses the source in place using the user’s authentication; federated modes for BigQuery, Spanner, Cloud SQL and AlloyDB are marked Preview.
Broad discovery and repeatable business reporting also differ. Google documents no table scoping or verified queries for federated relational-database queries. Curated conversational analytics agents instead support selected datasets and verified SQL, with additional setup. For a recurring business metric, our recommendation is to have the data owner define the permitted tables and accepted query before opening broader discovery.
Federation does not settle the destination of results. The connector security documentation says results return to the agent in the Gemini Enterprise app’s location and can be stored in conversations. Review source storage, returned rows, conversation storage and model processing separately.
The location matrix distinguishes storage residency from machine-learning processing. In-country access, including India, is allowlisted; enabling Gemini 3.8 Flash in unsupported in-country regions requires accepting routing to the global endpoint. A regional base app is therefore not enough evidence that every model and execution feature meets the same location requirement.
For the data-flow review, use the field-and-recipient mapping in Google’s Agentic Privacy Report: Map Data Flows Before Granting Access. Apply it to the actual connector and delegated task, not just the initial prompt.
Check whose authority reaches the final action
Ask the implementation team to trace a task from employee to coworker or subagent to the downstream tool. Record the identity used at each step, what it may read or change, who can revoke it, and which log records the action. A separate coworker account should not be assumed to inherit the employee’s permissions.
The Agent Gateway documentation describes outbound calls being denied without a matching access policy. Destinations can be registered resources or URLs explicitly covered by policy. It also distinguishes inbound from outbound controls: Gemini Enterprise supports the gateway’s outbound mode, while Agent Runtime supports both. Do not assume a single policy setting protects every direction or runtime.
Before granting write access, obtain feature-specific answers on approval steps, audit-log access, memory deletion and cross-model data handling. The reviewed launch material does not establish all of these operational details, and this guide does not claim to have tested the controls.
Price the whole workflow, and plan for a cap to stop it
No all-inclusive price for the new workplace-agent configuration was established by this research. In the documented subscription workflow, overages and spend controls require an invoiced billing account and at least one active, non-trial subscription. A trial alone cannot validate that production billing setup.
Model rates are only one line item. The published Gemini 3.8 Flash rate card lists introductory global standard-serving rates of US$0.75 per million uncached input tokens and US$3.75 per million output tokens through December 31, 2026, rising to US$1.50 and US$7.50 respectively on January 1, 2027. Output includes reasoning. The page notes that promotional pricing is delivered through credits; ask how those credits affect the quote. These are inference rates, not an agent subscription price.
Cloud Billing’s spend-cap rules are equally important: caps cover one project and one eligible service, and enforcement uses gross estimated costs without savings or credits. Enforcement is not instantaneous, in-flight requests can finish, persistent resources can keep accruing charges, and subscription costs are excluded. A cap is not a ceiling on the entire invoice.
The Gemini cost guide places app, Agent Platform and coding-tool spend under a shared project/service scope. Our operational inference is that shared spending can create a shared interruption point. Separate critical work from experimentation where appropriate, and name the person responsible for recovery. Manually lifting an enforced cap leaves it lifted for the rest of that budget period unless its amount is increased; full service recovery can take up to an hour.
What a useful pilot should produce
Start with one enabled workflow that produces an output a reviewer can check, such as a draft assembled from approved documents. Keep external actions out of the initial scope unless they are necessary and explicitly authorized. Before expansion, require:
A feature-and-access record: edition, region, models, connectors and any early-access dependency.
An authority-and-data map: permitted reads and writes, delegated identities, result storage and processing locations.
A separately reviewed comparison with the existing workflow: the same authorized inputs and acceptance rules, recorded configurations, timestamped outputs, failures, reviewer effort and all-in cost per accepted task.
Include subscription allocation, model usage, connector and query costs, runtime, storage, support and human review in that comparison. Expand when the workflow meets your acceptance criteria and operating constraints—not merely because the platform can reach more systems.
Methodology: AI-assisted reporting and analysis based on Google’s announcement, release notes, operational documentation and pricing, checked on October 8, 2026. No hands-on testing or independently controlled performance or cost comparison was conducted.