For two years, the first question in business AI has been which model to buy. Google’s answer at its Gemini at Work 2026 event, on October 8, was to ask a different question: which system should remember how your company works? The product it unveiled, the Gemini agent, is pitched in Thomas Kurian’s announcement as “your new single, universal agent for work.” It answers questions, handles knowledge work, makes images and writes and runs code. It also coordinates other agents and, notably, can route jobs to Anthropic’s Claude as well as to Google’s own models.
A disclosure first. This site’s drafting assistant is Claude, made by Anthropic, so a story in which Claude is a supported model is one where we have an interest. We have kept to Google’s own pages and to technology reporting, and we mark vendor claims as vendor claims.
A prompt box that keeps working after you close the laptop
The core idea is continuity. Google says the Gemini agent runs in the cloud with “a single set of memories, context, and one personalization graph,” so work that “takes hours or days keeps running after you close your laptop.” You can assign it a job, schedule a recurring one or have it react to an event. Access is meant to be everywhere: web, iOS and Android, Windows and Mac desktops, a command line, and inside Google Workspace, Microsoft 365 and Slack.
When a request has many steps, Gemini creates temporary, job-specific agents, each with its own identity, and coordinates parallel and sequential steps. Think of them as contractors hired for one assignment and released afterwards. That is a different thing from the second kind of agent Google describes, which is meant to stay.
Little of this is new in pieces. Google spent the year assembling the parts: long-running agents and a progress inbox in Gemini Enterprise, a separate Agent Platform for building and governing agents, and Workspace features that turn email threads into documents and decks. The October announcement is mostly about putting them behind one front door.
Coworkers with a badge, a calendar and a name in version history
The most concrete idea is the coworker agent. Rather than each employee having a private assistant, a team can describe a role, such as an event coordinator or a financial analyst, and Gemini builds an agent with a persistent job. In Workspace it gets what a new hire would: its own account with an email address, a calendar, Drive storage and a listing in the company directory.
Google’s example is a product launch. A marketing manager tags the agent in a document comment. It suggests an edit, replies in the thread and appears under its own name in the version history. Colleagues can see which changes were the agent’s, and which were not.
Two limits matter. The agent acts under its own identity, not yours, and Google says it sees only what you share with it, which means the existing sharing rules still decide what it can open. Analyst Keith Kirkpatrick of Futurum, in his assessment of the launch, calls these identity-bearing coworkers the most consequential part, since agent identity is becoming a baseline requirement across the category.
Where the memory comes from: tools, skills and four kinds of recall
An agent with an email address but no knowledge of the business would be an expensive intern. Google builds knowledge from three layers.
Tools connect Gemini to systems the company already runs. The announcement lists Confluence, Microsoft Office and Teams, Slack, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres and Snowflake, plus any server that speaks the Model Context Protocol, and an enterprise registry lets teams publish their own connectors. Skills are reusable instructions stored like modular prompts, for example how to assemble a monthly sales report, with a global library and a registry for sharing inside a company. Google also says Gemini learns from each run, improving consistency and saving token costs.
Context is the layer with the most interesting vocabulary. Google describes four types of memory: session memory for the task in front of the agent, even when it runs for days; semantic memory, a structured knowledge base built from documents, conversations and other agents; procedural memory for how a job is done, including skills the agent writes for itself; and episodic memory of everything it has done before. Google says Gemini “onboards itself the way a new hire would.”
The payoff Google shows is small but telling. Ask it to schedule a meeting with “the usual team of regional event leads next week” and it infers the attendees from chat groups and earlier threads, checks calendars and starts the email thread, without the user supplying a name. A second feature, built on Workspace Intelligence, notices when a manager emails a request for a slide deck and offers the recipient a one-click way to hand it to Gemini.
The net-margin problem, and why definitions matter more than prompts
The finance example in the announcement deserves attention because it shows what Google thinks the hard part is. Ask an AI for “net margin for the quarter” and the difficulty is not arithmetic. It is knowing which records count, and how this company defines the term.
Google’s answer pairs BigQuery with Knowledge Catalog, which maps business definitions and reads metrics from tools such as Databricks, dbt, LookML or SAP. Gemini identifies the right data, applies the company’s definition, and writes the query. The query can be saved, so the team can rerun the report on demand without paying for the model to regenerate it. For data engineers, new skills generate PySpark code, notebooks and models, and troubleshoot pipelines.
This is a quiet but revealing choice. Much of the value is in a company’s own definitions, not in the model, which is exactly the thing a platform wants to hold on to.
Google applies the same logic to regulated industries. Gemini for Financial Services, in preview, ships with more than 50 foundational skills and draws on providers such as FactSet, LSEG and S&P Global, along with SEC filings and internal data. Google says outputs carry confidence scores, explicit methodologies, data lineage and checkable citations, and names CME Group and Deutsche Bank. A legal version, also in preview, is said to inherit matter-level permissions and ethical walls from document systems such as NetDocuments and iManage, with Harvey and Onit as partners. Cooley is building an agent that redacts personal and confidential information before filings become public. Government, healthcare and retail are listed as coming soon.
Claude inside Google’s house
Now the part that gives the story its twist. Google says the agent orchestrates “across our Gemini family of models and Claude models from Anthropic today, and other leading private and open models in the future.” The reasoning is practical: “the best model for the task is not always the largest one.”
Smart Routing is how that plays out. It triages work so each job runs on the model that gives the best performance at the lowest cost, so a short summary does not need the same horsepower as a long software task. Android Headlines’ report says lightweight work goes to Flash-class models and heavier reasoning to models such as Claude, though Google’s own post describes routing in general terms and we found no official routing table.
Be careful with the word “today.” The announcement says Claude is supported, but reports on the launch describe third-party model selection as early access, and the full experience as limited to selected customers. We could not confirm a general-availability date in any official page.
From an analyst’s view, Kirkpatrick calls the Claude support “a pragmatic move” that removes a common reason for building a separate orchestration layer. That may be the sharpest reading of the news: Google is not conceding the model race, it is declining to make it the main event.
Guardrails: a firewall for agents and a switch for the budget
Letting software act across a company’s systems makes the controls the product. Google frames governance around four questions: who the agent is, what it may do, what it did, and what it should never touch.
On identity, Google says every agent has its own cryptographically attested identity, governed like an employee with least-privilege permissions. That identity is stamped into logs and into any virtual machine created to run code on the agent’s behalf. Access to outside systems uses identity propagation with standards such as OAuth, and every action is written to an audit trail attributed to the agent, not to a person. Agents run in an Agent Sandbox with its own network boundary, and an Agent Gateway acts as what Google calls an “AI network firewall,” with all traffic in, out and between agents passing through it. Its example policy is blunt: agents may not open documents classified “Need to Know,” written once and applied everywhere.
Cost is the other half. Google says per-token prices have fallen 98% since 2024, but enterprise volume has grown, which is the same tension we explored in why cheaper AI can still push compute demand up. Administrators can set a hard spending cap per project in the Cloud Billing Console. Gemini monitors token use and sandbox costs, and when the cap triggers, the project’s agent pauses until someone resumes it with a click. Spending is tracked per project so costs can be charged back to departments. The 98% figure and the savings claims are Google’s, and we have not verified them independently.
None of this removes the risks of agents acting on each other’s output. Our piece on how NVIDIA approaches containing AI agents covers the broader problem: a bad instruction can travel from one agent to the next. A gateway and a sandbox help, but a company still decides which permissions are sensible and when a human must approve.
The lock-in question nobody has answered yet
Here is the tension. If the model is swappable, the thing that is not swappable is everything around it: skills the agent wrote, memory it accumulated, the permissions it was granted and the history of what it did. Google says Gemini can work with data where it already lives, such as Salesforce or Microsoft 365, without requiring a move to Google. That is true of the data. It is a different matter for what the agent has learned about running your processes.
Google’s published pages describe skills, connectors and registries, but we found nothing explaining how a whole coworker, with its memory, permissions and responsibilities, could be exported to another platform. That is not proof it can’t be done. It means the answer is not in the announcement, and a buyer should ask before the agents become load-bearing.
Analyst Carmi Levy, quoted by Reworked from Computerworld, put the strategy in one line: Google is trying to become the gatekeeper of the enterprise agent “operating system,” powered by whichever model fits best. For a customer, that could be a feature or a dependency. Likely both.
Rivals are building the same thing from different doors
Google is not alone, and each rival starts from its own strength.
Microsoft’s September Copilot update adds Cowork for handing off whole tasks, and Autopilot, an agent that “can continue working on tasks independently, even when the user is away from their computer,” which Microsoft said would move into private preview at the end of that month. Microsoft’s advantage is the apps companies already use, and Futurum names Microsoft 365 Copilot as Google’s main rival.
OpenAI’s dots, launched on September 29, are always-on assistants powered by GPT-6 Astra, each with its own cloud computer. OpenAI says specialist dots will get their own identity, credentials and system access for defined roles, starting with focused enterprise pilots, and that it is working with Microsoft to bring them into Agent 365.
Anthropic’s Claude Tag, announced June 23, takes the Slack-first route: one shared Claude per channel, which builds context from that channel, can work asynchronously and keeps memories scoped to boundaries the administrators set. It is in beta for Enterprise and Team customers, with token spend limits and an action log. This is our own interest again, so read that description as Anthropic’s.
Note what the three share with Google: identity, persistent context, spend controls and logs. These are becoming the standard checklist. The competition is shifting from whose model scores higher to whose platform holds the company’s working knowledge.
What exists today and what is still a slide
It helps to separate announcement from availability. Google’s post does not give a general-availability date for the agent, and says the Financial Services and Legal versions are in preview. Reporting on the launch describes it as a private preview for select Gemini Enterprise customers. Capabilities such as multi-step background delegation, mobile and desktop access and third-party model choice are described as early access in the material we reviewed, so existing Gemini Enterprise customers should not assume everything shown on stage is switched on.
Price is the open question. Google told VentureBeat, as relayed by Reworked, that the agent will come built into Gemini Enterprise “at no additional charge” for eligible customers, and a Google representative told Futurum it does not intend to charge a per-seat licence for Workspace coworker agents. But how long-running jobs, sub-agents and sandbox time will be metered is not spelled out. Futurum flags that metering as the factor likely to decide whether coworker agents scale beyond pilots.
Evidence is thin too. Google’s post lists a long roll call of customers, with claims such as a bank cutting document review from an hour to five minutes. Those figures describe earlier Gemini Enterprise deployments and are vendor-reported, not results for the new agent. We found no independent test of how reliably it handles long jobs across several applications.
A sensible way to read the launch, then, is as a statement of direction with real components behind it. For a company evaluating it, the useful questions are plain ones. Which of these features are in my tenant now? What does a month of agent activity cost? Can I see and export what the agent knows? And if I changed the model underneath, would the work still hold? Google says yes to the last one. The other three are still open.
Editor’s note: This article was researched, drafted and revised with the help of Claude, an AI assistant made by Anthropic, working from official pages and the reporting linked in the text. Because this story involves Anthropic, we state that connection openly. Nicolae Panait, editor and owner of Webitlio, chose the topic, reviewed the facts and made the final editorial decisions.
