Google Cloud today presented the Gemini agent, a single agent for work that answers questions, produces documents and media, and writes and runs code from one prompt box, with Anthropic's Claude models available beside Google's own. According to Google Cloud chief executive Thomas Kurian, the system arrives with agent identities, a sandbox, a network gateway and project-level spend caps; according to Sharon Prosser, who leads SMB and Scaled at Google Cloud, small and midsize businesses are in early access now.

In Short

Google today showed one AI assistant for work that can answer questions, write documents, make images and write code, and that keeps working in the cloud after people close their laptops. It matters to companies of every size because the assistant can switch between Google's AI models and Anthropic's Claude models, and because a budget limit can stop it when spending reaches a set amount. Small businesses are in early access now, and Google says everyone else gets access "soon" without naming a date.

One agent, many surfaces

"Work now starts in the prompt window," Kurian wrote in a post adapted from his keynote at Gemini at Work 2026. According to Kurian, nearly 500 Google Cloud customers each processed more than one trillion tokens in the past year, nearly 80% of all Google Cloud customers use its AI products, and nearly 90% of the Fortune 100 use Gemini Enterprise. The figures are Google's own and have not been independently audited.

The agentic capabilities rest on six architectural principles listed in the post. The first, a unified agent, covers chat, autonomous work on assigned objectives and code generation through one interface. Omnipresent access spans the web, iOS and Android, Windows and Mac desktops, the command line, Google Workspace, Microsoft 365 and Slack; the agent can also be embedded in third-party applications, where it runs as a headless agent with no interface of its own.

Persistent execution moves the work into the cloud. The agent runs there and keeps one set of memories, context and a single personalization graph across devices, so work lasting hours or days continues after a laptop is closed. Under multi-agent orchestration, it can assemble a roster of sub-agents - temporary, job-specific agents, each with its own identity - to take on multi-step tasks in parallel or in sequence. It can also act as a coworker agent, a team member with a persistent role. According to Google, coworker agents have dedicated identities, including their own @agents.company.com email addresses, their own persistent storage, and access only to the context that people supply.

The remaining two principles are being deeply contextual, meaning the agent arrives knowing a company's tools, data and work history and learns from each interaction, and model choice. On the latter, Gemini is the agent and the model underneath is a separate decision. Google says the agent orchestrates across its Gemini family and Claude models from Anthropic today, with other private and open models to follow. The stated rationale is economic: the best model for a task is not always the largest, and matching model to job raises accuracy on hard work while lowering cost on simple work. Google cited the sportswear brand On as an early tester of the selection capability, and pointed to PayPal, which routes 10 million multi-model requests every week.

Tools, skills and memory

Three layers carry company-specific knowledge. A tools registry connects the agent to collaboration software such as Confluence, Teams and Slack; development tools including Git and Jira; enterprise platforms such as Salesforce and ServiceNow; and databases including BigQuery, Databricks, Postgres and Snowflake. It also works with any Model Context Protocol (MCP) server, inside or outside a company network, and teams can publish their own tools to a shared registry.

Skills are reusable instructions, stored as modular prompts, that teach the agent multi-step tasks. Gemini ships with a global library, departments can publish custom skills to a company registry, and individuals can write personal ones. Memory comes in four kinds: session memory for the task at hand, semantic memory built from documents and conversations, procedural memory for how a job gets done (including skills the agent writes for itself), and episodic memory of past work.

Inside Workspace

Within Workspace, Gemini operates in Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, carrying the same memory, skills and controls, according to Google. The post describes three modes: personal assistance, such as arranging a meeting with the usual regional event leads by reading chat-space membership and the previous event's thread; proactive delegation, in which Workspace Intelligence flags a manager's emailed request for a slide deck and offers a single-click handoff; and the coworker agent.

A coworker agent receives its own Workspace account - an email address, a calendar, a Drive and a place in the company directory - and colleagues add it to a Chat space or @mention it. In the post's example, a marketing manager asks an events coordinator agent to draft a launch readiness document; the agent can also suggest edits in a Doc and reply in comment threads, appearing under its own name in version history. It acts under its own identity, sees only what has been shared with it, and, according to Google, no outside connector holds the data.

Model choice meets the budget question

Kurian framed the commercial problem in two clauses. Two factors determine whether an enterprise agent program "succeeds or stalls", he wrote: whether it can be governed, and whether it can be afforded. Per-token prices have dropped 98% since 2024, according to Google Cloud, yet enterprise AI volume has "exploded". His formulation: "Running every simple loop through a premium model quickly breaks corporate budgets."

Google lists three controls. Multi-model orchestration picks the model per job. Smart Routing triages workloads so that each runs on the model delivering maximum performance at the lowest possible cost. Real-time spend caps set a hard limit on a project's AI spend in the Cloud Billing Console. Gemini monitors token usage and sandbox costs, and if a cap is triggered, that project's agent pauses; work can resume with a single click in the console. Because tracking is per project, companies can charge AI costs back to specific departments.

The mechanism has precedent inside Google. Project Spend Caps arrived in the Gemini API on March 16, 2026, a monthly dollar limit per project that took about ten minutes to take effect, after developers had reported billing errors in August 2025. The Cloud Billing version is described as tracking token usage and sandbox costs, but neither document states an enforcement delay or any dollar thresholds.

Other vendors have moved the same way. Cloudflare put dollar budgets on AI Gateway in open beta on June 5, 2026, blocking further requests by default once a limit is reached. The pressure behind such controls shows in survey data: KPMG found that 49% of 2,145 senior leaders had narrowed, delayed or paused AI agent deployments, with expected operating costs outrunning value as the main driver, and only 35% said their AI operating costs were fully visible and actively monitored.

Model prices keep moving underneath all of this. Anthropic priced Claude Haiku 5.5 at $0.10 per million input tokens for prompts up to 100,000 tokens on October 7, with availability on Google Cloud among other clouds; the figures are vendor-reported. Neither Google post names which Claude models the Gemini agent can route to, or says what the agent itself costs.

Governance: four questions

Kurian reduced agent governance to four questions: who the agent is, what it is permitted to do, what it did and where that can be seen, and what it is barred from touching. Identity, policy and observability answer the first three, he wrote; the fourth belongs to Agent Gateway.

Every agent gets its own identity, "cryptographically attested and governed like an employee", with least-privilege permissions. The identity is stamped into the logs that capture the agent's work and into any virtual machine launched to run code on its behalf. Permissions are role-based and approved by security administrators; when the agent connects to an external system, its identity is mapped and propagated through standards such as OAuth. Every action goes to an audit trail attributed to the agent rather than to a person, which security teams can watch in real time through observability tools.

At the network layer, all Gemini agents run tasks inside an Agent Sandbox with its own network boundary. All traffic - in, out and between agents - passes through Agent Gateway, which Google describes as "an AI network firewall" enforcing policy in real time. A policy is written once, the example being "agents may not open documents classified Need to Know", and then applies to every agent in the company.

The design speaks to problems that standards bodies are only beginning to list. MLCommons published a first draft cataloguing 25 privacy risk vectors for AI agents, including over-privileged data handshakes over protocols such as MCP and the lack of a standardized, privacy-preserving way to log agent activity across ecosystems. Neither Google post addresses how the audit trail would meet that second gap.

Data and analytics skills

Data and analytics is the first domain-specific skill set. For data and machine-learning engineers, Google says Gemini generates PySpark code, provides notebooks to edit and test it, trains models, and troubleshoots pipeline problems on its own. For business users, operational reporting skills tied to BigQuery and the Knowledge Catalog construct and save a query; teams can then run saved reports on demand "without incurring token costs".

Three components keep the answers grounded. The Knowledge Catalog maps business definitions such as "net margin" and "addressable market" once so that every agent uses them, and Gemini reads metrics where they sit, whether in Databricks, dbt, LookML or SAP. Smart Storage addresses unstructured data, which Google puts at 90% of the enterprise total; it enriches those objects in place and writes context back onto the object itself, so the data inherits its existing security posture. The Borderless Lakehouse lets Gemini query Amazon S3 and Azure Data Lake with no variable egress fees, read from Salesforce Data 360, SAP, ServiceNow and Workday without copying data, and federate Apache Iceberg tables across Databricks Unity, Snowflake Horizon and AWS Glue. Generated code runs on a Managed Spark Service with Lightning Engine or in BigQuery.

Customer figures accompany the section. Bloomberg Media lifted SQL query accuracy by 63% during initial development after grounding its data agents in the Knowledge Catalog, and its chief technology officer, William Anderson, said grounding AI "in a trusted institutional context" gives confidence in the accuracy and quality of every insight. Snap cut diagnostic troubleshooting from 30 minutes to 30 seconds, and Etsy reports data joins up to 60% faster after migrating three petabytes.

Industry packages and silicon

Industry specialization comes next. Gemini for Financial Services and for Legal are in preview, with Government, Healthcare and Retail listed as coming soon. The financial services version carries more than 50 foundational skills and draws on FactSet, LSEG, S&P Global, SEC filings and proprietary repositories, showing confidence scores, methodologies, data lineage and source citations; CME Group and Deutsche Bank are named users. The legal version inherits matter-level permissions and ethical walls from document management platforms such as NetDocuments and iManage.

Underneath sits hardware. Google says its TPU 8i system delivers 80% better price-performance than the prior generation, and it lists four model families: Argon for frontier reasoning, Flash for speed and volume, Omni for generative media, and Gemma for lightweight open-weights edge work.

The small-business track

Prosser's post addresses the segment directly. Millions of small and midsize businesses (SMBs) already use Google AI across Gemini Enterprise, Google Ads and Google Workspace, according to Google, and their usage of Google Cloud's AI tools increased more than fivefold year over year. The Gemini agent, she wrote, "is in early access with SMB customers today and will be available to all customers soon". The post gives no date, price or count of participants.

It names connectors to Asana, Box, Clay, Docusign, Dropbox, GitHub, LegalZoom, Notion, Salesforce, Shopify, Slack and Wix, and says the agent can take a defined role such as project manager, bookkeeper, inventory manager or marketing consultant. The same model orchestration, smart routing and project-level spend caps described for enterprises apply.

The customer examples in the SMB post concern Gemini Enterprise rather than the new agent. KLog.co, a Chilean logistics technology company, reports cutting manual data entry errors by more than 90% and a tenfold increase in document processing capacity. NEEOH, a Brazilian out-of-home advertising platform, uses the product to standardize secure AI use so teams can produce campaign copy and pitch proposals faster.

Customer claims, as reported by Google

Kurian's post carries dozens of customer examples. Bradesco, the Brazilian bank, reports cutting document review from one hour to five minutes while reducing risk inconsistencies by 60%. At Wesfarmers, an internal agent at Bunnings saved half a million hours of administrative work, and shopping agents raised conversion rates at Kmart and Officeworks by up to three times. Ulta Beauty reports a three-fold rise in digital sales conversion for its shopping assistant covering 30,000 products. The Home Depot cites four-fold faster phone resolution.

None of these entries states a baseline, a sample period or a measurement method, and all of them originate in Google's own blog. They describe Gemini Enterprise deployments; the Gemini agent itself is new.

Why the announcement matters to marketing teams

Neither post describes features specific to Google Ads. The advertising platform appears once, as one of the services SMBs already use, and "marketing consultant" appears as one of four example roles. The advertising relevance lies elsewhere, in the plumbing.

Google already ships assistants inside its advertising products: Ads Advisor and Analytics Advisor, Gemini-powered tools for Google Ads and Google Analytics, were presented in November 2025. The Gemini agent sits at a different layer, a cross-application worker that connects to systems of record through connectors and MCP servers. On that protocol, Google released an open-source, read-only MCP server for the Google Ads API on October 7, 2025, and Meta opened an ads MCP server and command-line interface to Claude and ChatGPT in an open beta on April 29, 2026. Usage is already visible: Ads Uploader counted 17.5% of 36,289 Meta ad batches in September as submitted by AI agents, although the data is vendor-supplied, unaudited and covers one month.

Access rules for such connections are also changing. Google's Ads developer policies, published on August 31, 2026, prohibit programmatic proxies, a category whose definition names MCP servers that only re-expose Google Ads capabilities. How the Gemini agent's connectors would interact with those rules is not stated in either post.

The cost side has an advertising analogue as well. Draft Digital replaced twelve direct MCP integrations with one buyer agent after team members hit Claude usage limits several times in a day, since each endpoint needed its own context load and the brief was restated on every branch. Kurian's post lists rosters of sub-agents and spend caps among the same features, though neither document gives figures for what such a roster consumes.

Several items remain unstated in the two posts: a price for the Gemini agent, a general availability date, the list of Claude models in use, the enforcement delay of the new caps, the number of businesses in early access, and any Google Ads integration. Each is a variable in how the product would reach advertisers and agencies.

Timeline

Summary

  • Who: Google Cloud, through chief executive Thomas Kurian and Sharon Prosser, vice president of SMB and Scaled; Anthropic's Claude models as a supported option; enterprise customers and small and midsize businesses as users.
  • What: The Gemini agent, a single agent for work with sub-agents and coworker agents, persistent cloud execution, four memory types, a tools and skills registry, model choice across Gemini and Claude, agent identity, an Agent Sandbox, Agent Gateway and project-level spend caps that pause an agent when a limit is reached.
  • When: Today. SMBs are in early access now; Google says wider availability will follow "soon" and gives no date. Financial Services and Legal packages are in preview, with Government, Healthcare and Retail still to come.
  • Where: Across the web, iOS and Android, Windows and Mac, the command line, Google Workspace, Microsoft 365 and Slack, announced at Gemini at Work 2026 and on the Google Cloud blog.
  • Why: Google frames the product around governance and cost, the two factors it says decide whether enterprise agent programs succeed, with per-token prices down 98% since 2024 but agent volume straining budgets. For advertisers and agencies, the open questions are pricing, availability, Claude model coverage and how connectors interact with platform rules for MCP access.