Google Cloud Introduces Gemini, a Universal Work Agent

Google announced one agent for work on October 8

Google Cloud introduced the Gemini agent at Gemini at Work 2026 on October 8, 2026. The company positions it as a single enterprise agent that can answer questions, complete knowledge work, generate media, analyze data, and write and run code. The promise is outcome delegation from one prompt window rather than assembling a separate agent for each application.

The primary company announcement is dated October 8. Google Cloud’s much longer keynote adaptation displays October 9 even though it was already publicly retrievable during this review on October 8. We therefore use October 8 as the announcement date and preserve the one-day source inconsistency rather than silently normalizing it.

Persistent and multi-agent execution are the central change

Google says Gemini can answer interactively, run scheduled tasks, respond to events and keep long jobs running in the cloud after a user closes a device. It describes one memory and personalization graph across sessions and surfaces, with session, semantic, procedural and episodic memory. Teams can also create projects that constrain the context, skills and tools used for a body of work.

For multi-step objectives, Google says Gemini can create temporary sub-agents and coordinate parallel or sequential work. It also describes persistent coworker agents with dedicated identities, storage and @agents.company.com email addresses. Those are consequential authority and lifecycle changes: a proof of chat quality does not establish safe operation for an agent that persists, receives events or acts under its own organizational identity.

It connects to work systems and routes across model families

The launch names Google Workspace, Microsoft 365, Slack, Confluence, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, Snowflake and desktop files among the agent’s reachable systems. Google also describes an enterprise tool registry, a shared skills registry and support for Model Context Protocol servers. The product FAQ says more than 100 integrations are supported, with access following each user’s existing permissions.

Google says the agent can choose among Gemini and Anthropic Claude models today, with additional private and open models planned. Smart Routing is meant to select a model for quality and cost. That is a vendor-described orchestration policy, not an independently verified optimizer. Buyers need logs that expose the selected model, version, route, tokens, tools, failures and charge for each production task.

Several headline capabilities are still early access

The live product page says customers can contact sales or try Gemini Enterprise Plus at no cost for 30 days. It also carries a broad warning: not every feature shown is available now. Google specifically identifies multi-step background delegation, Gemini on mobile and desktop, and third-party model choice as early access, with availability to all customers promised later.

That caveat is more decision-useful than the breadth of the launch demo. Availability must be recorded per edition, account, region, surface and feature. The reviewed pages do not provide a universal activation date, a launch-specific API identifier or a matrix showing which announced memory, coworker, connector, data and industry capabilities are enabled for each customer today.

Pricing is sales-led and the API contract is incomplete

Google advertises the 30-day Gemini Enterprise Plus trial and contact-sales routes, but the reviewed launch pages do not publish a launch-specific seat price, token tariff, connector charge, sandbox price or minimum commitment. The keynote says Gemini works through a single API and describes real-time project spend caps, yet it does not name a Gemini-agent API endpoint, request schema, version, quota or service-level objective.

Do not substitute underlying Gemini or Claude model prices for the total agent price. A routed task may include model calls, retrieval, connectors, tool execution, sandbox compute, storage and retries. Obtain a written quote and meter a representative trial. The operational unit should be cost per accepted completed task, with failed attempts and human correction included.

Governance claims need configured evidence

Google says each agent receives a cryptographically attested identity, fine-grained role-based permissions and an attributable audit trail. It describes OAuth identity propagation, isolated Agent Sandbox execution, an Agent Gateway that applies network policy, and project-level spend caps that pause work when a limit is reached. Its product FAQ says business prompts, outputs and tuning data are not used to train Google models or models for other customers.

These are first-party product and policy statements, not an AccessAllGPT security assessment. Existing user permissions can still be too broad, connected content can contain hostile instructions, and model routing changes the processor and behavior behind a task. Test least privilege, approval boundaries, connector revocation, identity attribution, policy denial, data lineage, export, deletion, incident response and budget-stop behavior on the exact purchased configuration.

What enterprise teams should do next

Start with one reversible workflow whose source systems, actions and acceptance criteria are already understood. Ask Google for an edition-and-region capability sheet, model-routing controls, data and retention terms, connector boundaries, API documentation, audit event schema, support commitments, service levels and complete commercial terms. Mark every demo-only or early-access capability as unavailable until it appears in the target account.

Trial when the necessary features are enabled and the contract is reviewable. Constrain the agent to read-only or approval-gated actions while measuring completion quality, latency, routing, connector errors, interventions and full task cost. Wait when a required feature is early access or unpriced. Reject deployment when identity, audit, data, spend or rollback controls cannot be demonstrated under production-shaped tests.

Disclosures

AccessAllGPT received no Google credentials, trial access, credits, quote, briefing, benchmark data, review, payment or compensation for this article. Google and the cited customers did not sponsor, review or endorse it. AccessAllGPT did not use, test, score or rank the Gemini agent. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Google, Anthropic or organizations cited. Publication-wide relationships are listed on the disclosures page.